今年2月ごろから始めた論文斜め読みが千本を超えたので、リストを掲載。
分野は、物体認識、Deep Learningの軽量化、Neural Architecture Searchがメイン。
適当な掲載方法が見つからず体裁が悪いのだが、とりあえず上げておく。
Year | Affiliation | Title | Category | Key word | Comment | Performance | Prior | Link | OSS | Related info. | |
---|---|---|---|---|---|---|---|---|---|---|---|
2019 | Xiaomi | Fair DARTS: Eliminating Unfair Advantages in Differentiable Architecture Search | NAS | unfair advantage in an exclusive competition zero-one loss | dominant operationsの可視化が参考になる P-DARTS が高速高精度 | 3GPUdays 75.6%@CIFAR-100 | P-DARTS | https://arxiv.org/pdf/1911.12126.pdf | |||
2019 | Carnegie Mellon University | DEEP MULTIVARIATE MIXTURE OF GAUSSIANS FOR OBJECT DETECTION UNDER OCCLUSION | Object Detection | occlusion | mAP:62.1/50.9/38.4(occ3段階) | soft-NMS, DeepVoting | https://arxiv.org/pdf/1911.10614.pdf | ||||
2019 | Xiamen University, Huawei | Multinomial Distribution Learning for Effective Neural Architecture Search | NAS | MdeNAS, Multinomial Distribution Learning | 0.16GPUdays | DARTS | https://arxiv.org/pdf/1905.07529.pdf | https://github.com/tanglang96/MDENAS | |||
2019 | Tianjin university | Exploiting Operation Importance for Differentiable Neural Architecture Search | NAS | EoiNAS | DARTS拡張、MdeNASの方が優れている | 0.6GPUdays | MdeNAS | https://arxiv.org/pdf/1911.10511.pdf | |||
2019 | Bosch | Meta-Learning of Neural Architectures for Few-Shot Learning | NAS, meta-learning | meta loss | gradient-based meta-learningとNASを初めて組み合わせた | AutoMeta | https://arxiv.org/pdf/1911.11090.pdf | ||||
2019 | Simon Fraser University | BA-Net: Dense Bundle Adjustment Network | Bundle Adjustment | depth parameterization | 下記からの引用 end-to-endでBA camera motionとdepthの同時推定 | 95.21 ms(320 × 240解像度2枚) | https://arxiv.org/pdf/1806.04807.pdf | ||||
2019 | AIST | 論文ではないが単眼距離推定のサーベイ | https://speakerdeck.com/kanezaki/ji-jie-xue-xi-woyong-ita3ci-yuan-detaren-shi-nituite?slide=61 | ||||||||
2019 | Chemnitz University of Technology, Southampton Solent University | Traffic Signs Recognition and Distance Estimation using a Monocular Camera | TSR | YOLO+ active learning 学習データの比率は、posi:45%/unsure:35%/nega:20% BBox面積で距離を推定 | https://www.researchgate.net/profile/Shadi_Saleh3/publication/337362850_Traffic_Signs_Recognition_and_Distance_Estimation_using_a_Monocular_Camera/links/5dd3f04b299bf11ec8600590/Traffic-Signs-Recognition-and-Distance-Estimation-using-a-Monocular-Camera.pdf | ||||||
2019 | Search to Distill: Pearls are Everywhere but not the Eyes | Knowledge Distillation | Architecture-aware Knowledge Distillation (AKD) | RL | NASNet | https://arxiv.org/pdf/1911.09074.pdf | |||||
2019 | Huawei, The Hong Kong University of Science and Technology | MULTI-OBJECTIVE NEURAL ARCHITECTURE SEARCH VIA PREDICTIVE NETWORK PERFORMANCE OPTIMIZATION | NAS | graph convolutional network (GCN), Bayesian Optimization | random searchに対し+3% DARTSと同等(CIFAR-10) | DARTS, ASNG-NAS | https://arxiv.org/pdf/1911.09336.pdf | ||||
2003 | Massachusetts Institute of Technology | Contextual Priming for Object Detection | Object Detection | Context-Driven Focus of Attention | 下記からGIST featureの引用 手で設計したattention+saliency map | http://people.csail.mit.edu/torralba/IJCVobj.pdf | |||||
2018 | Indiana University Bloomington | Toddler-Inspired Visual Object Learning | Object Detection | attention, GIST featuresz(extracted GIST features [31] from each object instance and computed pairwise GIST distances (L2 norm) across all instances ) | object variabilityをGIST distanceで評価 | http://vision.soic.indiana.edu/papers/diversity2018nips.pdf | |||||
2019 | Detection of Microaneurysms in Fundus Images Based on an Attention Mechanism | Object Detection | image quality equalization(Otsu+CLAHE) Feature Fusion Based on Attention | 前処理でコントラストを上げる score mapとbinary GTの差のloss(attention) | AP:0.757 | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6827155/pdf/genes-10-00817.pdf | |||||
2016 | Stanford University | Image quality object detection accuracy | Object Detection | 画質低下の影響 | http://cs231n.stanford.edu/reports/2016/pdfs/287_Report.pdf | ||||||
2018 | UBTECH Sydney | Deep Ordinal Regression Network for Monocular Depth Estimation | Monocular Depth Estimation | DORN | KITTI4位(code公開ではトップ) DeepLabに近い | absErrorRel:8.78%(500ms) | https://arxiv.org/pdf/1806.02446.pdf | https://github.com/hufu6371/DORN | |||
2019 | Huawei | RefinedMPL: Refined Monocular PseudoLiDAR for 3D Object Detection in Autonomous Driving | Object Detection | Approximate Foreground Separation (AFgS), Distance-Stratified Sampler (DSS), Supervised Sparsification | monocular 3D 教師あり・なしの組合せ | AM3D | https://arxiv.org/pdf/1911.09712.pdf | ||||
2019 | DeepMind, Google | Fast Sparse ConvNets | Acceleration | XNNPACK, sparse matrix-dense matrix multiplication (SpMM) | Weight sparsityでは高速化しないという誤解を解く動機 | 1.3 − 2.4×高速化 | https://arxiv.org/pdf/1911.09723.pdf | ||||
2019 | Huawei | SM-NAS: Structural-to-Modular Neural Architecture Search for Object Detection | NAS, Object Detection | Backbone-Neck (fusion)-RPN-Headの構造は変えない | Titan-VでCOCO mAP~44.7@50ms2.5 | https://arxiv.org/pdf/1911.09929.pdf | |||||
2019 | Google Research | EfficientDet: Scalable and Efficient Object Detection | Object Detection | EfficientDet, BiFPN | 複数レベルを通るパスの追加 | Titan-VでCOCO mAP~44.7@50ms | https://arxiv.org/pdf/1911.09070.pdf | ||||
2019 | Beihang University | Learning Spatial Fusion for Single-Shot Object Detection | Object Detection | YOLOv3 + ASFF | Pyramid level間のfusion | Tesla V100でCOCO mAP43.9%@34ms | https://arxiv.org/pdf/1911.09516.pdf | https://github.com/ruinmessi/ASFF | |||
2019 | Microsoft | Label-similarity Curriculum Learning | Image Classification | Curriculum learning probability distribution over classes as target label | 動的にラベルを変える 蒸留で似たことやってる例なかったか? | https://arxiv.org/pdf/1911.06902.pdf | |||||
2019 | University of Illinois at Urbana-Champaign, XMotors.ai | NAIS: Neural Architecture and Implementation Search and its Applications in Autonomous Driving | NAS | DNN/implementation co-design, FPGA-oriented Bundle generation | FPGAの動作fpsを制約とする | YOLOv3の6倍以上高速 | https://arxiv.org/pdf/1911.07446.pdf | ||||
2019 | Cornell University, Fudan University, Tsinghua University, Facebook AI Research | Multi-Scale Dense Networks for Resource Efficient Image Classification | Image Classification | MSDNet | 下記でDynamic inferenceのSOTAとされた論文 | DenseNet+3% | https://arxiv.org/pdf/1703.09844.pdf | ||||
2019 | Peking University | S2DNAS: Transforming Static CNN Model for Dynamic Inference via Neural Architecture Search | NAS | dynamic inference(入力に応じて動的に計算グラフを切り替え高速化), Channel-wise | チャネル分割の仕方を探索 MSDNetのような階層型分岐構造の探索 | 57%パラメータ削減で精度-0.7%(MobileNetV2, CIFAR-100) | https://arxiv.org/pdf/1911.07033.pdf | ||||
2019 | Shandong University | Lane Recognition Algorithm Using the Hough Transform Based on Complicated Conditions | Lane Detection | Dark-Light-Dark-threshold | ラインの左右両端の微分のminが閾値以上かで判定 直線・障害物/カーブなしでしか評価していない | https://www.scirp.org/pdf/jcc_2019111914261439.pdf | |||||
2019 | Alibaba, Northeastern University | DARB: A Density-Adaptive Regular-Block Pruning for Deep Neural Networks | Model Compression | block-max weight masking (BMWM) irregular pruning | Acc.+3%, 1/25圧縮 | https://arxiv.org/pdf/1911.08020.pdf | |||||
2019 | Baidu | 2nd Place Solution in Google AI Open Images Object Detection Track 2019 | Object Detection | Top-k Voting-nms | 複数モデルのensemble | https://arxiv.org/pdf/1911.07171.pdf | |||||
2016 | University of Washington | XGBoost: A Scalable Tree Boosting System | Boosting | いわずと知れたXGBoostの元論文 | https://arxiv.org/pdf/1603.02754.pdf | https://lib-arts.hatenablog.com/entry/tree_based_algo1 | |||||
2017 | Indiana University, Pacific Northwest National Laboratory | A Survey of Methods for Collective Communication Optimization and Tuning | |||||||||
2005 | Argonne National Laboratory | Optimization of Collective Communication Operations in MPICH | collective communication algorithms "All reduce", "All gather" and "Reduce scatter" | LightGBM公式で’LightGBM implements state-of-art algorithms[9].’として引用されている。 以下Wikipedia引用”MPICH(以前はMPICH2と呼ばれていた)は、自由に利用可能なポータブルなMPI実装である。MPIとは、並列計算で使われる分散メモリアプリケーションのためのメッセージパッシングの標準の一つである。” | https://github.com/microsoft/LightGBM/blob/master/docs/Features.rst#references https://www8.cs.umu.se/kurser/5DV050/VT11/F2.pdf | ||||||
2016 | Peking University, Microsoft Research | A Communication-Efficient Parallel Algorithm for Decision Tree | Parallel Voting Decision Tree (PV-Tree) | LightGBM論文の引用 LightGBMのtree構造の工夫を理解するために重要 | http://papers.nips.cc/paper/6381-a-communication-efficient-parallel-algorithm-for-decision-tree.pdf | ||||||
2013 | Graduate University of Chinese Academy of Sciences | Justifying the importance of color cues in object detection: a case study on pedestrian | Object Detection | 物体検出におけるLUVの優位性 歩行者検出FPPI評価でベスト | https://www.researchgate.net/profile/Lei_Qin6/publication/267721938_Justifying_the_Importance_of_Color_Cues_in_Object_Detection_A_Case_Study_on_Pedestrian/links/549534780cf29b9448211619.pdf | ||||||
2019 | Texas A&M University | Fourier Spectrum Discrepancies in Deep Network Generated Images | GAN | GANの生成画像は高周波成分が過剰に大きい傾向 | https://arxiv.org/pdf/1911.06465.pdf | ||||||
2019 | University of California | TinyCNN: A Tiny Modular CNN Accelerator for Embedded FPGA | 16-bit fixed-point data in a CNN accelerator (FPGA) | 3% accuracy loss | https://arxiv.org/pdf/1911.06777.pdf | ||||||
2019 | Volvo | Automated Augmentation with Reinforcement Learning and GANs for Robust Identification of Traffic Signs using Front Camera Images | TSR, DA | YOLOv3, bounding box GAN (BBGAN) | DAの効果は基本的に小さい模様 | 昼は効果なく、夜がPrecision向上大きい(precision/recall from 0.70/0.66 to 0.83/0.71 for nighttime images) | https://arxiv.org/pdf/1911.06486.pdf | ||||
2019 | Chinese Academy of Sciences | DATA: Differentiable ArchiTecture Approximation | NAS | Ensemble Gumbel-Softmax (EGS) | 1GPUdaysでAmoebaNet-Cなみ | https://papers.nips.cc/paper/8374-data-differentiable-architecture-approximation.pdf | https://github.com/XinbangZhang/DATA-NAS | ||||
2019 | Sandia National Laboratories | RAPDARTS: Resource-Aware Progressive Differentiable Architecture Search | co-attention and co-excitation (CoAE) framework | 12GPUdays, DSO-NAS(1GPUdays)と同等精度 | https://arxiv.org/pdf/1911.05704.pdf | ||||||
2019 | National Tsing Hua University | One-Shot Object Detection with Co-Attention and Co-Excitation | https://papers.nips.cc/paper/8540-one-shot-object-detection-with-co-attention-and-co-excitation.pdf | https://github.com/timy90022/One-Shot-Object-Detection. | |||||||
2019 | University of Washington | Traffic Sign Detection and Recognition for Autonomous Driving in Virtual Simulation Environment | TSR | RetinaNet | https://arxiv.org/pdf/1911.05626.pdf | ||||||
2019 | University of Cambridge | Focused Quantization for Sparse CNNs | Model Compression | Focused compression (5 bits, sparse) | x16.3 | https://papers.nips.cc/paper/8796-focused-quantization-for-sparse-cnns.pdf | https://github.com/deep-fry/mayo | ||||
2019 | Learning from a Teacher using Unlabeled Data | Distillation | Blind Distillation+Fine-Tune with Labeled Data | https://arxiv.org/pdf/1911.05275.pdf | |||||||
2019 | Deep learning model for object detection | Unsupervisedでpose/depthを出す | https://aip.scitation.org/doi/pdf/10.1063/1.5133483 | ||||||||
2019 | KAIST | SimVODIS: Simultaneous Visual Odometry, Object Detection, and Instance Segmentation | VO | Simultaneous Visual Odometry, Object Detection, and Instance Segmentation (SimVODIS) utilizes image reconstruction from different camera views as a supervision signal | Mask-RCNN+FPN | https://arxiv.org/pdf/1911.05939.pdf | |||||
2019 | UCLA | PerspectiveNet: 3D Object Detection from a Single RGB Image via Perspective Points | Object Detection | 2D Box Branch, Perspective Points Branch, 3D Box Branch | SUN RGB-D | https://papers.nips.cc/paper/9093-perspectivenet-3d-object-detection-from-a-single-rgb-image-via-perspective-points.pdf | |||||
2019 | Chinese Academy of Sciences, Megvii | DetNAS: Backbone Search for Object Detection | NAS | FPN and RetinaNet Shuffle Unit | random baselineに対し<+1% | http://papers.nips.cc/paper/8890-detnas-backbone-search-for-object-detection.pdf | https://github.com/megvii-model/DetNAS. | ||||
2019 | Chongqing University of Posts and Telecommunications | A Multibranch Object Detection Method for Traffic Scenes | Object Detection | SSDより高速高精度 P-R評価 | 30ms(KITTI) | http://downloads.hindawi.com/journals/cin/2019/3679203.pdf | |||||
2019 | Telkom University | Design and implementation of obstacles detection in selfdriving car prototype | Object Detection | HSV Segmentation The distance formulas are produced by Geogebra Application with power trendline. | 面積から距離推定? | https://iopscience.iop.org/article/10.1088/1742-6596/1367/1/012020/pdf | |||||
2019 | University of Waterloo | EFFICACY OF PIXEL-LEVEL OOD DETECTION FOR SEMANTIC SEGMENTATION | Semantic Segmentation | out of distribution (OOD) detection | https://arxiv.org/pdf/1911.02897.pdf | ||||||
2019 | Dalian University of Technology | SFE-SSD: Shallow Feature Enhancement SSD for Small Object Detection | Object Detection | SSD, Feature Fusion | mAP+1.2%(Conv4 3 deconv + fusion, VOC2007) 85 FPS w/ 1080Ti | http://jmre.dlut.edu.cn/en/ch/reader/create_pdf.aspx?file_no=20190616&year_id=2019&quarter_id=6&falg=1 | |||||
2019 | Arizona State University | Structural Pruning in Deep Neural Networks: A Small-World Approach | Model Compression | hierarchically trimming Small-World model | intrinsic network propertyに注目 | Acc;0.17%, -66%param.(VGG) | https://arxiv.org/pdf/1911.04453.pdf | ||||
2007 | University of Waikato | Best-first Decision Tree Learning | Light GBMが参照しているleaf-wise growth best-firstと表現している その意味はimpurityの低減の最大化 | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.149.2862&rep=rep1&type=pdf | |||||||
2019 | A PROGRAMMABLE APPROACH TO MODEL COMPRESSION | Model Compression | https://arxiv.org/pdf/1911.02497.pdf | ||||||||
2019 | FRD-CNN: Object detection based on small-scale convolutional neural networks and feature reuse | Object Detection | fire module | mAP+2.2%(VOC2007) | MobileNet+SSD | https://www.nature.com/articles/s41598-019-52580-0.pdf | |||||
2019 | Huazhong University of Science and Technology | Localization-aware Channel Pruning for Object Detection | Model Compression | -75%で~2%drop | https://arxiv.org/pdf/1911.02237.pdf | ||||||
2019 | Carnegie Mellon University | SCL: Towards Accurate Domain Adaptive Object Detection via Gradient Detach Based Stacked Complementary Losses | Object Detection | stacked complementary losses (SCL) method domain adaptive | Domain AdaptationをFaster RCNNに応用し、霧シーン(FoggyCityscapes)の認識率向上 中間層からcontextを抽出するsub-networkを、勾配伝搬を一方通行にするdetach strategyを用いて学習 | https://arxiv.org/pdf/1911.02559.pdf | https://github.com/harsh-99/SCL | ||||
2019 | KU Leuven | A Systematic Analysis of a Context Aware Deep Learning Architecture for Object Detection | Object Detection | Contextual heatmap | Keras Mask R-CNNベース | memory network, Gated Bi-Directional (GBD) units | http://ceur-ws.org/Vol-2491/paper90.pdf | https://github.com/kbardool/Contextual-Inference-V2 | |||
2019 | Northern Border University | Pedestrian Detection for Advanced Driver Assistance Systems using Deep Learning Algorithms | Object Detection | The "Fire" module has a squeeze convolution layer | YOLO v2+SqueezeNet | mAP:75.8%, 32.4fps | SSD+SqueezeNet, Mask R-CNN | https://www.researchgate.net/profile/Yahia_Said/publication/337050186_Pedestrian_Detection_for_Advanced_Driver_Assistance_Systems_using_Deep_Learning_Algorithms/links/5dc2a0a34585151435ecba04/Pedestrian-Detection-for-Advanced-Driver-Assistance-Systems-using-Deep-Learning-Algorithms.pdf | |||
2017 | Microsoft Research | Resource-efficient Machine Learning in 2 KB RAM for the Internet of Things | Bonsai single, shallow, sparse tree sparsely projecting into a low-dimensional space | 下記引用 2kBメモリのDT NeuralNetより高精度 データセットが画像でないので性能がわからない | https://www.microsoft.com/en-us/research/uploads/prod/2017/06/kumar17.pdf | https://github.com/BonsaiAI https://docs.bons.ai/references/library-reference.html | |||||
2019 | LG Electronics | On-Device Machine Learning: An Algorithms and Learning Theory Perspective | edgeでの学習 | https://arxiv.org/pdf/1911.00623.pdf | |||||||
2019 | CEA, LIST | LapNet : Automatic Balanced Loss and Optimal Assignment for Real-Time Dense Object Detection | Object Detection | Per-Object Normalized Overlap | OSSなし ThunderNetより低速高精度 | mAP:38.2(COCO) ~80ms,M2Detに劣る(~55ms;DarkNet-53backboneではmAP37.6でぴったり一致) CenterNetには劣る | RetinaNet, FCOS | https://arxiv.org/pdf/1911.01149.pdf | |||
2019 | Aggregated Residual Dilation-Based Feature Pyramid Network for Object Detection | 改善効果は小 | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8840842 | ||||||||
2019 | Shanghai Jiao Tong University | Object Guided External Memory Network for Video Object Detection | Object Detection | selectively store valuable features | フレーム間整合性をattention融合/固定で? | http://openaccess.thecvf.com/content_ICCV_2019/papers/Deng_Object_Guided_External_Memory_Network_for_Video_Object_Detection_ICCV_2019_paper.pdf | |||||
2019 | National University of Defense Technology, Megvii Inc. | ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices | Object Detection | Spatial Attention Module | とにかく速い YOLOv3と同等 | CPU at 32.3 fps mAP:28.1(COCO) | Pelee, Tiny-DSOD | http://openaccess.thecvf.com/content_ICCV_2019/papers/Qin_ThunderNet_Towards_Real-Time_Generic_Object_Detection_on_Mobile_Devices_ICCV_2019_paper.pdf | https://github.com/mohhao/TF-Keras-ThunderNet | ||
2019 | University of Engineering and Technology | Classification of Degraded Traffic Signs Using Flexible Mixture Model and Transfer Learning | TSR | 特徴量次元に対しあるところで誤差はサチる この時130ms HOG-LBPでも向上 FC後にDR | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8866733 | ||||||
2019 | University of Wisconsin Madison | Resource Constrained Neural Network Architecture Search: Will a Submodularity Assumption Help? | NAS | RCAS, Lazy Cost Effective Greedy algorithm (LCEG) | 9GPUdays 2GPUdaysでDARTSと同等精度 DARTSよりTop-1 Accuracy+1.5% | MNasNet-92 | http://openaccess.thecvf.com/content_ICCV_2019/papers/Xiong_Resource_Constrained_Neural_Network_Architecture_Search_Will_a_Submodularity_Assumption_ICCV_2019_paper.pdf | ||||
2019 | South China University of Technology, Tencent | NAT: Neural Architecture Transformer for Accurate and Compact Architectures | NAS | Markov decision process (MDP), graph convolution network (GCN), policy gradient | あまりに微々たる改善 | https://arxiv.org/pdf/1910.14488.pdf | |||||
2018 | Texas A&M University | Improved Techniques for Learning to Dehaze and Beyond: A Collective Study | Object Detection | deghazeの効果はたいしてないという結論 | https://arxiv.org/pdf/1807.00202.pdf | ||||||
2019 | Shenzhen University | Does deep learning always outperform simple linear regression in optical imaging? | Compressed sensing | Linear Regression | プロジェクタ+1画素センサで実験、DLを上回る | https://arxiv.org/ftp/arxiv/papers/1911/1911.00353.pdf | |||||
2019 | Huawei | Auto-FPN Automatic Network Architecture Adaptation for Object Detection Beyond Classification | NAS, Object detection | Resource constraints | 4~16GPUdays SSD-ResNet101と大差なし(mAP:31.8(+0.6%)) | SSD | http://openaccess.thecvf.com/content_ICCV_2019/papers/Xu_Auto-FPN_Automatic_Network_Architecture_Adaptation_for_Object_Detection_Beyond_Classification_ICCV_2019_paper.pdf | ||||
2019 | Toshiba | SCALABLE DEEP NEURAL NETWORKS VIA LOWRANK MATRIX FACTORIZATION | Model Compression | complexity-based criterion SVD low-rank network | 学習後にmodel sizeを変更する model size変更時はBNのパラメータを修正すべき | CIFAR100で50%までは精度低下なし | slimmable networks | https://arxiv.org/pdf/1910.13141.pdf | |||
2019 | Tsinghua University, Northeastern University, … | Adversarial Robustness vs. Model Compression, or Both? | Model Compression | Concurrent Adversarial Training | Adv attackにロバストにするとcapacityが増える課題に対しpruning pruningはirregular/regularに分かれる | http://openaccess.thecvf.com/content_ICCV_2019/papers/Ye_Adversarial_Robustness_vs._Model_Compression_or_Both_ICCV_2019_paper.pdf | https://github.com/yeshaokai/ Robustness-Aware-Pruning-ADMM | ||||
2019 | Intel AI Lab | Neural Network Distiller: A Python Package For DNN Compression Research | Model Compression | PyTorchで軽量化を開発するPython OSS Compressionの代表技術のまとめとしても有用 | https://arxiv.org/pdf/1910.12232.pdf | https://github.com/NervanaSystems/distiller | |||||
2019 | Shadow Detection and Removal for Traffic Surveillance System | Shadow removal | HSV Shadow Detection Condition =F_V < avg_V && F_V > avg_V/2 && F_S < max_S/3 | 2枚入力 | https://s3.amazonaws.com/academia.edu.documents/60491667/284_Shadow_Detection_and_Removal_for_Traffic_Surveillance_System20190904-12736-lv1cw3.pdf?response-content-disposition=inline%3B%20filename%3DShadow_Detection_and_Removal_for_Traffic.pdf&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIWOWYYGZ2Y53UL3A%2F20191104%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20191104T064921Z&X-Amz-Expires=3600&X-Amz-SignedHeaders=host&X-Amz-Signature=f9449cb541d876855e03bc4983f47c9bf1f3c7aeb5c540336ad3bebb86e6a123 | ||||||
2019 | Shadow Detection and Removal Techniques: A Perspective View | Shadow removal | https://s3.amazonaws.com/academia.edu.documents/59927707/334_Shadow_Detection_and_Removal_Techniques_A_Perspective_View20190704-66578-h7fu65.pdf?response-content-disposition=inline%3B%20filename%3DShadow_Detection_and_Removal_Techniques.pdf&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIWOWYYGZ2Y53UL3A%2F20191104%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20191104T064930Z&X-Amz-Expires=3600&X-Amz-SignedHeaders=host&X-Amz-Signature=f54d721b15883e220960b2680da14ebd162012f92a55e7aa86efe220b5f89452 | ||||||||
2019 | LMU Munich | Tunability: Importance of Hyperparameters of Machine Learning Algorithms | HPO | tunability, XGBoost, AUC, accuracy and Brier score | http://www.jmlr.org/papers/volume20/18-444/18-444.pdf | ||||||
2009 | Multi-class AdaBoost | Boosting | SAMME forward stagewise additive modeling algorithm that minimizes a novel exponential loss | AdaboostのSAMME sklearnのensemble/_weight_boosting.py/_compute_proba_from_decisionに実装 | http://ww.web.stanford.edu/~hastie/Papers/SII-2-3-A8-Zhu.pdf https://web.stanford.edu/~hastie/Papers/samme.pdf | ||||||
2019 | RealityEngines.AI, CMU | BANANAS: Bayesian Optimization with Neural Architectures for Neural Architecture Search | NAS | Neural BayesOpt | DARTSと変わらず、より遅い | https://arxiv.org/pdf/1910.11858.pdf | |||||
2019 | Korea University | Reverse and Boundary Attention Network for Road Segmentation | road | free space 影・輝度飽和に頑健 端部の精度向上 | F1:96.30% 160ms(KITTI) | http://openaccess.thecvf.com/content_ICCVW_2019/papers/CVRSUAD/Sun_Reverse_and_Boundary_Attention_Network_for_Road_Segmentation_ICCVW_2019_paper.pdf | |||||
2019 | An Annotation Saved is an Annotation Earned: Using Fully Synthetic Training for Object Detection | Object Detection | 3D background model, synthetic training data 3D CAD Model+3D Pose Curriculum Curriculum strategy significantly outperforms random pose selection | 照明方向/照明色/noise/blurを変化させてrendering 追試困難 | 実写とぴったり同じ精度(Faster R-CNN) | http://openaccess.thecvf.com/content_ICCVW_2019/papers/R6D/Hinterstoisser_An_Annotation_Saved_is_an_Annotation_Earned_Using_Fully_Synthetic_ICCVW_2019_paper.pdf | |||||
2019 | Imperial College London | Efficient Structured Pruning and Architecture Searching for Group Convolution | Model Compression | GConv pruning Pointwise/Permutation-based channel permutations layerwise heuristic pruning | 73.77%,err+5.4%(CIFAR100) | CondenseNet | http://openaccess.thecvf.com/content_ICCVW_2019/papers/NeurArch/Zhao_Efficient_Structured_Pruning_and_Architecture_Searching_for_Group_Convolution_ICCVW_2019_paper.pdf | https://github.com/kumasento/gconv-prune | |||
2019 | Hana Institute of Technology | Efficient Decoupled Neural Architecture Search by Structure and Operation Sampling | NAS | EDNAS | DARTSと同精度 | 3.6GPUdays | DSO-NAS | https://arxiv.org/pdf/1910.10397.pdf | |||
2015 | 京都大学 | 今日からできるスパースモデリング | http://www-adsys.sys.i.kyoto-u.ac.jp/mohzeki/Presentation/lecturenote20150901.pdf | ||||||||
2019 | Sharif University of Technology, INRIA | A Comparative Study of Neural Network Compression | Model Compression | Optimal Brain Damage | https://arxiv.org/pdf/1910.11144.pdf | ||||||
2019 | Skolkovo Institute of Science and Technology | Automated Multi-Stage Compression of Neural Networks | Model Compression | Low-rank tensor approximations, MUSCO | 1.39×, mAP+2.0(FASTER R-CNN) | http://openaccess.thecvf.com/content_ICCVW_2019/papers/LPCV/Gusak_Automated_Multi-Stage_Compression_of_Neural_Networks_ICCVW_2019_paper.pdf | |||||
2012 | Katholieke Universiteit Leuven | Pedestrian detection at 100 frames per second | Object Detection | ACF論文でcascadeの手法として引用 VeryFast Soft cascade GPU | https://rodrigob.github.io/documents/2012_cvpr_pedestrian_detection_at_100_frames_per_second.pdf | ||||||
2017 | University of California, Davis, Google Research | GPU-acceleration for Large-scale Tree Boosting | Boosting | LightGBM We replace the ConstructHistogram function in LightGBM with our GPU implementation | GTX1080で2~4x高速化(building feature histograms) | https://arxiv.org/pdf/1706.08359.pdf | |||||
2019 | Zhejiang University, Singapore Management University, Alibaba-Zhejiang Universit | CSID: Center, Scale, Identity and Density-aware Pedestrian Detection in a Crowd | Object Detection | Identity and Density Map (ID-Map) ID-NMS | anchor-free pedestrian detectionのNMSの改良 | 160ms, missrate:5.8~46.6(CityPersons, 1080Ti) | https://arxiv.org/pdf/1910.09188.pdf | ||||
2019 | The University of Adelaide | FCOS: Fully Convolutional One-Stage Object Detection | Object Detection | centerness | 下記からの引用 anchor-free lossにcenterness(eq.3)を加えるのがポイント NMSは使う | RetinaNet | https://arxiv.org/pdf/1904.01355.pdf | ||||
2019 | Tohoku University | Analysis and a Solution of Momentarily Missed Detection for Anchor-based Object Detectors | Object Detection | soft thresholding SSD, M2Det | 東北大岡谷先生 課題:SSD/M2Detにおいて、フレームによって未検出になる 原因:物体の移動に対し、隣接する2つのanchor(default) boxの両方のスコアが下がるため未検出が発生する。従来の学習方法は、positive/negative sampleの判定が binary thresholding のためにIoUの変化に過敏であった。言い換えると、物体のScaling/Shifting/Aspect変化(anchor box境界近傍の変化)に対するロバスト性を考慮していなかった。 対策:IoU閾値近傍の物体に連続値のスコアを与えるsoft-thresholding(Fig.7参照)を適用した学習データ生成を行い、隣接2boxのスコアが同時に下がる現象を緩和。 | https://arxiv.org/pdf/1910.09212.pdf | |||||
2019 | MIT, Cisco | NASIB: Neural Architecture Search withIn Budget | NAS | augmenting the search space of ENAS Policy based Sampling from Search Space | 探索範囲12倍 探索効率がENASの4倍? | 1.5GPUdays | https://arxiv.org/pdf/1910.08665.pdf | ||||
2019 | Beijing Institute of Technology, Intel | SPCDet: Enhancing Object Detection with Combined Feature Fusing | Object Detection | Serial-parallel Combined Fusing, feature pyramid, Context fusing | 38.0ms@512x512 | RefineDet512 | http://proceedings.mlr.press/v101/wang19e/wang19e.pdf | ||||
2019 | FZI Research Center for Information Technology | Realtime 3D Object Detection for Automated Driving Using Stereo Vision and Semantic Information | Object Detection | Disparity Image+Semantic Map, Grid Map | KITTI | https://www.mrt.kit.edu/z/publ/download/2019/Koenigshof2019Objects.pdf | |||||
2019 | Astyx GmbH | Deep Learning Based 3D Object Detection for Automotive Radar and Camera | Object Detection | 下記引用 | https://www.astyx.com/fileadmin/redakteur/dokumente/Deep_Learning_Based_3D_Object_Detection_for_Automotive_Radar_and_Camera.PDF | ||||||
2019 | Astyx GmbH | Automotive Radar Dataset for Deep Learning Based 3D Object Detection | Object Detection | Semi-Automatic Labeling via Active Learning | radar dataset 公開はされていない | https://www.astyx.com/fileadmin/redakteur/dokumente/Automotive_Radar_Dataset_for_Deep_learning_Based_3D_Object_Detection.PDF | |||||
2018 | UC Berkeley,Adobe | Toward Multimodal Image-to-Image Translation | GAN | BicycleGAN LPIPS Distance | 下記からの引用 LPIP Distanceで画像多様性定量化 “image distance measure diversity”で検索してトップに来る | https://arxiv.org/pdf/1711.11586.pdf | https://github.com/junyanz/BicycleGAN | ||||
2018 | UC Berkeley, OpenAI, Adobe | The Unreasonable Effectiveness of Deep Features as a Perceptual Metric | LPIPS | image diversityの定量化 下記からの引用 | https://arxiv.org/pdf/1801.03924.pdf | https://www.github.com/richzhang/PerceptualSimilarity. | |||||
2019 | Universidad Aut´onoma de Barcelona | Controlling biases and diversity in diverse image-to-image translation | average the LPIPS distance between 19 random pairs of outputs for 100 different input images | https://arxiv.org/pdf/1907.09754.pdf | |||||||
2015 | A boosted multi-task model for pedestrian detection with occlusion handling | Multi-Task ACF Model, three decision stumps | ACFのocclusion対策 occlusion levelの異なるDTの組み合わせ | miss rate:35.81~82.37 (heavy以外はtop) | https://www.aaai.org/ocs/index.php/AAAI/AAAI15/paper/download/9879/9825 | ||||||
2006 | Inria | Coloring Local Feature Extraction | 下記からの引用 SIFTのcolor版 | https://lear.inrialpes.fr/people/vandeweijer/papers/eccv2006.pdf | |||||||
2012 | Color Attributes for Object Detection | Object Detection | ColorRobust hue descriptor (HUE) descriptors, Opponent derivative descriptor (OPP), Color names (CN), late fusion | 色特徴量/ACF拡張参考、color/shapeのfusionはlateの方がいい | http://cat.uab.es/~joost/papers/cvpr2012.pdf | ||||||
2018 | Huazhong University of Science and Technology | Pedestrian-Synthesis-GAN: Generating Pedestrian Data in Real Scene and Beyond | Data Augmentation | PS-GAN, Pix2Pix | 下記からの引用 | https://arxiv.org/pdf/1804.02047.pdf | |||||
2019 | University of Michigan, Google Cloud AI | Generative Modeling for Small-Data Object Detection | Object Detection | generative modeling, Pedestrian Detection | 少数データからの学習 | https://arxiv.org/pdf/1910.07169.pdf | |||||
2019 | Shandong University | A Small Traffic Sign Detection Algorithm Based on Modified SSD | TSR | SSD300 | mAP:0.85(CCTSDB dataset) | https://iopscience.iop.org/article/10.1088/1757-899X/646/1/012006/pdf | |||||
2011 | NICTA | Large Scale Sign Detection using HOG Feature Variants | TSR | Single Bin HOG Feature | HOGの変形 | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.727.207&rep=rep1&type=pdf | |||||
2016 | Graz University of Technology | Grid Loss: Detecting Occluded Faces | Object Detection | https://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Documents/lrs/pubs/opitz_eccv_16.pdf | |||||||
2019 | Benem´erita Universidad Aut´onoma de Puebla | Speed Bump Detection on Roads using Artificial Vision | road | stereo | bump検出 | https://www.polibits.cidetec.ipn.mx/2019_148_9/Speed%20Bump%20Detection%20on%20Roads%20using%20Artificial%20Vision.pdf | |||||
2019 | Baidu, University of Technology Sydney | One-Shot Neural Architecture Search via Self-Evaluated Template Network | NAS | SETN Gradient | 特筆すべき性能ではない | 1.8GPU days | GDAS | https://arxiv.org/pdf/1910.05733.pdf | |||
2019 | University of Waterloo | State of Compact Architecture Search For Deep Neural Networks | Model Compression | Generative Synthesis [9] 8 | https://arxiv.org/pdf/1910.06466.pdf | ||||||
2019 | Tianjin University | Mask-Guided Attention Network for Occluded Pedestrian Detection | Object Detection | Mask-Guided Attention VGG16+RoIAlign+MGA+FC CityPersons [31] and Caltech [7]. CityPersons [31] is a challenging dataset | FPPI評価 | Bi-Box | https://arxiv.org/pdf/1910.06160.pdf | ||||
2019 | Valeo, Cairo University, IRI BarcelonaTech | FuseMODNet: Real-Time Camera and LiDAR based Moving Object Detection for robust low-light Autonomous Driving | Object Detection | Moving Object Detection FlowNet RGB + rgbFlow + lidarFlow | DARK-KITTIデータセットを構築(MOD用) | 18fps mIOU:75.3(KITTI) | https://arxiv.org/pdf/1910.05395.pdf | https://sites.google.com/view/fusemodnet | |||
2019 | Academia Sinica | IMMVP: An Efficient Daytime and Nighttime On-Road Object Detector | Object Detection | ResNet-18 Backbone Cascade R-CNN subnet BDD100K | mAP:0.460(IoU @ 0.5) 57MB, 510ms, 56 GOPs, 11M params 24.3 FPS@FHD | https://arxiv.org/pdf/1910.06573.pdf | |||||
2019 | University of California, Berkeley | Improving a Quality of 3D Object Detection by Spatial Transformation Mechanism | Object Detection | endpoint box regression module(epBRM) | 点群にBox Regression | AVOD, F-PointNet, PointPillars - | https://arxiv.org/pdf/1910.04853.pdf | ||||
2019 | University of Rochester | Learning Sparsity and Quantization Jointly and Automatically for Neural Network Compression via Constrained Optimization | Model Compression | Support pruning Support quantization End-to-end optimization | 4.41%dropで26.7× (ImageNet) 0%dropで836×(CIFAR10,ResNet-50) | CLIP-Q | https://arxiv.org/pdf/1910.05897.pdf | ||||
2019 | Latent AI | Bit Efficient Quantization for Deep Neural Networks | Model Compression | MobileNet-SSD | 7-bitまでは精度が落ちない | lossなしに52%減 | https://arxiv.org/ftp/arxiv/papers/1910/1910.04877.pdf | ||||
2017 | Mahidol University International College, IBM Research AI | Low-Latency Sliding-Window Aggregation in Worst-Case Constant Time | Aggregation (e.g., computing the sum or geometric mean) | 下記引用 | http://hirzels.com/martin/papers/debs17-daba.pdf | ||||||
2018 | Mahidol University International College, IBM Research AI | Sliding-Window Aggregation Algorithms | http://hirzels.com/martin/papers/encyc18-sliding-window.pdf | ||||||||
2013 | The University of Electro-Communications | An Efficient and Scalable Implementation of Sliding-Window Aggregate Operator on FPGA | https://core.ac.uk/download/pdf/147692472.pdf | ||||||||
2019 | IBM Research | Benchmarking and Optimization of Gradient Boosting Decision Tree Algorithms | Boosting | XGBoost, LightGBM and Catboost Hyper-Parameter Optimization, Bayesian Optimization, GPyOpt | https://arxiv.org/pdf/1809.04559.pdf | ||||||
2019 | KU Leuve | Fast Gradient Boosting Decision Trees with Bit-Level Data Structures | Boosting | Gradient boosting decision tree BitBoost does not (yet) support multi-class classification | 精度を維持した学習の高速化 2クラスのみ | speed up model construction 2 to 10 times | https://ecmlpkdd2019.org/downloads/paper/557.pdf | https://github.com/laudv/bitboost | |||
2011 | Brno University of Technology | Acceleration of Object Detection using Classifiers | Object Detection | user defined cost function, WaldBoost | https://pdfs.semanticscholar.org/091e/02bd7f6bc827b581b3b0c90a34b93b35ee0b.pdf | ||||||
2016 | Universitat Autonoma de Barcelona | GPU-based pedestrian detection for autonomous driving | Object Detection | 歩行者検出 | HOG+LBPで単品よりもmissrate-10% GPUでx8高速化 | https://www.sciencedirect.com/science/article/pii/S1877050916309395 | |||||
2017 | An Up-to-Date Comparison of State-of-the-Art Classification Algorithms | Image Classification | 分類のサーベイ 速度ではGBDT(XG-Boost)がベスト | https://www.researchgate.net/publication/315955430_An_Up-to-Date_Comparison_of_State-of-the-Art_Classification_Algorithms | |||||||
2019 | Wuhan University | Context-Aware Convolutional Neural Network for Object Detection in VHR Remote Sensing Imagery | Object Detection | VGG4,5 blockからcontext抽出 | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8846087 | ||||||
2019 | Johannes Kepler University Linz | Patch Refinement - Localized 3D Object Detection | Object Detection | LiDAR only, BEV | https://arxiv.org/pdf/1910.04093.pdf | ||||||
2013 | Restoring An Image Taken Through a Window Covered with Dirt or Rain | Dirt Detection | CNN, patch-level network | 汚れ検知&修復 | real-timeは今後の課題 | BM3D | http://openaccess.thecvf.com/content_iccv_2013/papers/Eigen_Restoring_an_Image_2013_ICCV_paper.pdf | ||||
2011 | Lens Shading Correction for Dirt Detection | Dirt Detection | HE, Morphological Operation, B-spline | https://www.csie.ntu.edu.tw/~fuh/personal/LensShadingCorrectionforDirtDetection.pdf | |||||||
2014 | Detection of Camera Artifacts from Camera Images | Dirt Detection | thresholded NCC map 100 NCC maps accumulated | フレーム間NCC | https://www.honda-ri.de/pubs/pdf/1000.pdf | ||||||
2016 | Dirt Detection on Camera Module Using Stripe-Wise Background Modeling | Dirt Detection | https://www.researchgate.net/profile/Bin_Liao/publication/311008159_Dirt_Detection_on_Camera_Module_Using_Stripe-Wise_Background_Modeling/links/5a5d287e458515c03ede81da/Dirt-Detection-on-Camera-Module-Using-Stripe-Wise-Background-Modeling.pdf | ||||||||
2019 | Georgia Institute of Technology | Challenging Environments for Traffic Sign Detection: Reliability Assessment under Inclement Conditions | TSR | 画像劣化耐性評価 | a precision of 0.55 and a recall of 0.32, F0.5 score of 0.48 and F2 score of 0.35 | https://arxiv.org/pdf/1902.06857.pdf | https://github.com/olivesgatech/CURE-TSD | ||||
2019 | Yandex, Moscow Institute of Physics and Technology | CatBoost: unbiased boosting with categorical features | CatBoost | Pythonパッケージあり Light GBMより高精度 | Light GBMと同じ速度 | Light GBM | https://arxiv.org/pdf/1706.09516.pdf | https://github.com/catboost/catboost | https://catboost.ai/docs/concepts/python-installation.html | ||
2013 | York University | 50 Years of Object Recognition: Directions Forward | Object Detection | https://www.researchgate.net/publication/257484936_50_Years_of_object_recognition_Directions_forward | |||||||
2019 | PLA Army Engineering University, | Real-Time Object Detection in Remote Sensing Images Based on Visual Perception and Memory Reasoning | Object Detection | self-attention pre-screening | リモセン固有でBBoxの角度も出す | Recall:98.82, 14.03ms | DSSOD, As-Net | http://scholar.google.co.jp/scholar_url?url=https://www.mdpi.com/2079-9292/8/10/1151/pdf&hl=ja&sa=X&d=16238942437737322476&scisig=AAGBfm1YySSzv19FCHvlCPQJYTUBNv2OVQ&nossl=1&oi=scholaralrt&hist=5RIGpeIAAAAJ:13208294427371959306:AAGBfm3UqH6LPjfgfMRfQL8OPgtOHOZJ9A | |||
2009 | Silesian University of Technology | Error Analysis of Stereo Calibration and Reconstruction | Camera Calibration | SVD-based stereovision algorithms | 6以上の対応点座標の行列MをSVD 視差が小さいと誤差が拡大 誤差敏感度と再構成誤差は必ずしも対応しない | https://www.researchgate.net/publication/221055026_Error_Analysis_of_Stereo_Calibration_and_Reconstruction | |||||
2016 | Grid Loss: Detecting Occluded Faces | Object Detection | grid loss | CNNのocclusion対策 | https://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Documents/lrs/pubs/opitz_eccv_16.pdf | ||||||
2018 | United International University | MEBoost: Mixing Estimators with Boosting for Imbalanced Data Classification | Object Detection | RUSBoost | https://arxiv.org/pdf/1712.06658.pdf | ||||||
2017 | Japan Automobile Research Institute | Edge filter for road white-line detection using brightness gradient approximation by discrete values | Lane Detection | brightness-gradient approximation non-linear filter that approximates a low-pass filter plus differential calculus | 遅いし、FP-FNトレードオフが解決できているわけでもない | 西日でF-measure0.739 | https://www.jstage.jst.go.jp/article/jamdsm/11/2/11_2017jamdsm0029/_pdf/-char/ja | ||||
2018 | Indiana University–Purdue University Indianapolis | Road Edge Detection in All Weather and Illumination via Driving Video Mining | Lane Detection | ML天候分類 | https://www.researchgate.net/publication/329980954_Road_Edge_Detection_in_All_Weather_and_Illumination_via_Driving_Video_Mining | ||||||
2007 | University of Missouri, POSTECH | Coherent Line Drawing | Line Detection | http://umsl.edu/mathcs/about/People/Faculty/HenryKang/coon.pdf | |||||||
2005 | Optical Flow: Techniques and Applications | https://www.dgp.toronto.edu/~donovan/stabilization/opticalflow.pdf | |||||||||
2012 | Idiap Research Institute | Exact Acceleration of Linear Object Detectors | Object Detection | Patchworks マルチスケール画像を1枚にタイリングしFourier変換で効率的処理 FFTW+Eigen Discriminatively Trained Deformable Part Models | HOGの時間が主要になった | x6.4>高速化と精度向上 | https://publications.idiap.ch/downloads/papers/2012/Dubout_ECCV_2012.pdf | http://www.idiap.ch/scientific-research/resources. | |||
2017 | Nanyang Technological University | Fast and Accurate Pedestrian Detection using Dual-Stage Group Cost-Sensitive RealBoost with Vector Form Filters | Object Detection | 1)Feature Approximation 2)selective classification 3)selective scale processing Dual-Stage Group Cost-sensitive RealBoost scale of support Filtered Channel Features (FCF) framework+vector form filters ”the detection process only on alternate scales of the pyramid”具体的な記述なし | negative samples with complex backgroundに注目 1)coarse to fineで分類 2)ACFの補完は精度低下することを指摘。検出用のスケールを限定する方針。 ACFの後継が、LDCF[27], InformedHaar [40], Checkerboards [42] and RotatedFilters[41] cost-insensitive(AdaBoost)の後にcost-sensitive boostingを行う2段階 | 処理時間ACFの3倍、miss rate-15.14% | https://chengjuzhou.bitbucket.io/paper/ACMMM2017/ACMMM2017.pdf | ||||
2014 | Universidade Federal de Minas Gerais | An Optimized Sliding Window Approach to Pedestrian Detection | Object Detection | Sliding Windows Random Filtering | regressionで位置合わせ | 1.4% of detection windows is enough to detect 83% of the pedestrians on the INRIA dataset | http://www.ssig.dcc.ufmg.br/wp-content/uploads/2014/09/2014-An-Optimized-Sliding-Window-Approach-to-Pedestrian-Detection.pdf | ||||
2015 | Chinese Academy of Sciences | Convolutional Channel Features | Object Detection | Power law | ACFの代替としてCNN+boosting CCFに10 HOG+LUVを追加して改善 画像ピラミッド高速化のために、Patchwork(1チャネル化、FCありCNNにのみ恩恵、1.5 ∼ 3 times speed up) ACFの近傍レベルのfeatureをscaleする方法を転用 | ACFに対しMiss Rate -9.44%(Caltech pedestrian benchmark [14]) 1.5 ∼ 3 times speed up add 16 pixel padding | ACF, LDCF | https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/Yang_Convolutional_Channel_Features_ICCV_2015_paper.pdf | https://bitbucket.org/binyangderek/ccf/src/master/ | ||
2019 | Brown University, Sensetime | IMPROVING ONE-SHOT NAS BY SUPPRESSING THE POSTERIOR FADING | NAS | Posterior Convergent NAS (PC-NAS) weight sharing oneshot NAS のposterior fading problemへの対策 (true/proxy parameter posteriorの距離がsupergraphのモデル数に従い増大) | MobileNetV2に対しtop-1 +2.1% ResNet50/101の中間の性能をどっちよりも軽く(latency 3/4) | ProxylessNAS-gpu, EfficientNet-B0 | https://arxiv.org/pdf/1910.02543.pdf | ||||
2019 | Duke University | Deep Learning for Case-Based Reasoning through Prototypes: A Neural Network that Explains Its Predictions | Interpretability | autoencoder | 下記からの引用 中間表現の分類器を作る? | Fashion-MNIST | https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/download/17082/16552 | ||||
2019 | UT Austin | Splitting Steepest Descent for Growing Neural Architectures | Training technique | Progressive Training | 下記著者 | https://arxiv.org/pdf/1910.02366.pdf | |||||
2019 | EPFL | Overcoming Multi-model Forgetting | NAS | Weight Plasticity Loss (WPL) | weight-sharingの課題 複数モデルで共有する重みが、一方の学習の更新で損なわれる ENASに組み合わせた | NAO | https://arxiv.org/pdf/1902.08232.pdf | ||||
2019 | UT Austin, Facebook | ENERGY-AWARE NEURAL ARCHITECTURE OPTIMIZATION WITH FAST SPLITTING STEEPEST DESCENT | NAS | continuous splitting process Splitting one neuron into two off-springs Rayleigh-Quotient Gradient Descent speed up the splitting process of Liu et al. (2019b) | weight-sharingの課題:multi-model forgetting problem 演算コストのloss追加 PyTorch実装 | MobilenetV2に対しtop-1 +1.1%, 同程度MAC (ImageNet) | AMC | https://arxiv.org/pdf/1910.03103.pdf | https://github.com/dilinwang820/fast-energy-aware-splitting | ||
2019 | The State Key Laboratory of Novel Software Technology | Soware Engineering Practice in the Development of Deep Learning Applications | 技術者へのインタビュー 設計フェーズでの性能評価・構造設計が課題 66%がTF使ってる バグ探しは、ツール・コードレビュー・ログ・ブレークポイント・敵対的サンプル quality of data(QoD) が重視されてきている | https://arxiv.org/pdf/1910.03156.pdf | |||||||
2019 | University of Chicago | Band-limited Training and Inference for Convolutional Neural Networks | Model Compression | band-limited training can effectively control the resource usage | 下記からの引用 ノイズ耐性を上げて軽量化 Fourier係数の一定数を削除 | 精度は必ず低下 | http://proceedings.mlr.press/v97/dziedzic19a/dziedzic19a.pdf | ||||
2019 | University of California, Oregon State University | BENCHMARKING NEURAL NETWORK ROBUSTNESS TO COMMON CORRUPTIONS AND PERTURBATIONS | Data Augmentation | Gaussian data augmentation and adversarial training Our IMAGENET-C dataset consists of 15 types of algorithmically generated corruptions | 下記からの引用 バンドパスフィルタ | https://arxiv.org/pdf/1903.12261.pdf | |||||
2019 | Google, UC Berkeley | A Fourier Perspective on Model Robustness in Computer Vision | Training technique | Gaussian data augmentation, adversarial training | DAの限界は特定の劣化種別に特化してしまうこと Gaussian data augmentation, adversarial trainingともにハイパスの劣化には強いが、ローパスの劣化には弱いことを検証 このトレードオフに対するAutoAugmentの有効性を確かめた Fourier解析をした点が新しい Fourier heat map:精度の各周波数敏感度 | AutoAugment | https://arxiv.org/pdf/1906.08988.pdf | ||||
2019 | Volvo, Halmstad University | Deep Neural Network Compression for Image Classification and Object Detection | Object Detection | network agnostic model compression dynamical clustering approach YOLOv3, FasterRCNN | pruning→quantizationの2段階 | pruned about 95%, 110× less memory(CIFAR10) YOLOは60%で精度も下がってるのでダメ | https://arxiv.org/pdf/1910.02747.pdf | https://github.com/AhrazA/modelcompression-2019 | |||
2019 | University of Canterbury | xYOLO: A Model For Real-Time Object Detection In Humanoid Soccer On Low-End Hardware | Object Detection | Tiny YOLOより10~100倍高速、1/40サイズ | https://arxiv.org/pdf/1910.03159.pdf | ||||||
2019 | UC Berkeley, Microsoft Research Asia | DEFORMABLE KERNELS: ADAPTING EFFECTIVE RECEPTIVE FIELDS FOR OBJECT DEFORMATION | Deformable Kernels, effective receptive field (ERF) experts of Conditional Convolutions have better correlation with object semantics than their scales | 可視化の点でも有用(T-SNE, support) deformable conv.の発展形 global(GAP)+local(conv) | Conditional Convolutions | https://arxiv.org/pdf/1910.02940.pdf | |||||
2019 | AutoAugment: Learning Augmentation Strategies from Data | Data Augmentation | AutoAugment ShearX/Y, TranslateX/Y, Rotate, AutoContrast, Invert, Equalize, Solarize, Posterize, Contrast, Color, Brightness, Sharpness, Cutout [12], Sample Pairing [24] Proximal Policy Optimization (PPO) [53]を適用したが、GAやrandom searchでさらに改善するかもという記載。 | PILで実装 RLで最適なDAを探す NAS論文と同じcontrollerとchildが登場 | Cutout | http://openaccess.thecvf.com/content_CVPR_2019/papers/Cubuk_AutoAugment_Learning_Augmentation_Strategies_From_Data_CVPR_2019_paper.pdf | https://github.com/tensorflow/models/tree/master/research/autoaugment. | ||||
2017 | Microsoft, Peking University | LightGBM: A Highly Efficient Gradient Boosting Decision Tree | Image Classification | LightGBM | https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree.pdf | ||||||
2019 | ONE-SHOT NEURAL ARCHITECTURE SEARCH VIA COMPRESSIVE SENSING | NAS | Fourier analysis of Boolean functions AutoAugmentを適用 | DARTSのDAGと同じ, architecture encode randomly sampled sub-graph edgesのみBPで更新 構造をvalidation performanceに変換する関数を | 0.25GPU days | DARTS | https://openreview.net/pdf?id=B1lsXREYvr | ||||
2017 | Integer Occupancy Grids : a probabilistic multi-sensor fusion framework for embedded perception | OGM | OGMが基礎から詳しい | https://tel.archives-ouvertes.fr/tel-01680375/document | |||||||
2019 | Federal University of Parana (UFPR) | ANDA: A Novel Data Augmentation Technique Applied to Salient Object Detection | Data Augmentation | 物体の背景を入れ替えるDA | https://arxiv.org/pdf/1910.01256.pdf | ||||||
2019 | Universidade Federal do Espírito Santo | Bio-Inspired Foveated Technique for Augmented-Range Vehicle Detection Using Deep Neural Networks | Object Detection | path planに基づきROI設定 | https://arxiv.org/ftp/arxiv/papers/1910/1910.00944.pdf | ||||||
2019 | University of Michigan | LiTE: Light-field Transparency Estimation for Refractive Object Localization | pose estimation, Angular Filter, 3D EPI Filters, Unreal Engine 4 | https://arxiv.org/pdf/1910.00721.pdf | |||||||
2019 | Carnegie Mellon University | Towards Unifying Neural Architecture Space Exploration and Generalization | NAS, Model Compression | NN-Mass, Network Science-based NAS, Generalization | 構造から汎化性能を予測するための理論的基礎 lottery ticket hypothesisは構造を視野に入れていないことを指摘 | https://arxiv.org/pdf/1910.00780.pdf | |||||
2019 | University of Waterloo | YOLO Nano: a Highly Compact You Only Look Once Convolutional Neural Network for Object Detection | Object Detection | Residual Projection-Expansion-Projection Macroarchitecture Fully-connected Attention Macroarchitecture | モデルサイズは1/15だが、演算量が40%減でしかない。 Tiny YOLOに対する精度向上が大きい | ∼26.9 FPS and ∼48.2 FPS (At 15W and 30W power budgets, Jetson AGX Xavier) 4.57Bops | https://arxiv.org/pdf/1910.01271.pdf | ||||
2019 | Tsinghua University | Global Sparse Momentum SGD for Pruning Very Deep Neural Networks | Model Compression | lottery tickets, momentum-SGD-based optimization on-the-fly pruning learning-based pruning | x5 Compress Ratio | https://arxiv.org/pdf/1909.12778.pdf | |||||
2019 | Model Pruning Enables Efficient Federated Learning on Edge Devices | Model Compression | Raspberry Pi | Sample Basedで精度低下がなだらか(でも下がるので興味ない) | https://arxiv.org/pdf/1909.12326.pdf | ||||||
2019 | Bosch | Automated design of error-resilient and hardware-efficient deep neural networks | NAS | error resilience at the algorithm level Hardware-focused neural architecture design LEMONADE | GTSRB | https://arxiv.org/pdf/1909.13844.pdf | |||||
2019 | RandAugment: Practical data augmentation with no separate search | Data Augmentation | RandAugment | Python code 簡単に精度上げられそうで注目 | COCOでmAP+1.3 | https://arxiv.org/pdf/1909.13719.pdf | |||||
2019 | Recognition of Various Objects from a Certain Categorical Set in Real Time Using Deep Convolutional Neural Networks | Object Detection | SSD, Raspberry Pi 3b board | http://ysip3.computational-logic.org/proceedings/paper_5.pdf | |||||||
2019 | Google Brain, Baidu | EPNAS: Efficient Progressive Neural Architecture Search - Supplementary Materials | NAS | Resource constraints | https://bmvc2019.org/wp-content/uploads/papers/0456-supplementary.pdf | ||||||
2019 | Huawei | STACNAS: TOWARDS STABLE AND CONSISTENT OPTIMIZATION FOR DIFFERENTIABLE NEURAL ARCHITECTURE SEARCH | NAS | CIFAR10で1GPUdays | https://arxiv.org/pdf/1909.11926.pdf | ||||||
2019 | Data augmentation with Symbolic-to-Real Image Translation GANs for Traffic Sign Recognition | TSR | Pix2pix | GANによるDA Contrastによる従来DAが最高精度(不思議) Blur, Brightness, Contrast, Displacement, Occlusion, Rotation, and Scalingがベースライン | https://arxiv.org/pdf/1907.12902v1.pdf | ||||||
2019 | ROAD DETECTION USING TRAFFIC SIGN INFORMATION | 特許、標識位置を手掛かりに路面推定 | https://patentimages.storage.googleapis.com/1e/ff/61/2fe93e2280f957/US20190272435A1.pdf | ||||||||
2017 | University of South Carolina | Detecting Small Signs from Large Images | TSR | SSD+VGG-16 SmallObject-Sensitive-CNN (SOS-CNN) | 遅いらしい(employed a sliding window strategy, it is time consuming) | Hierarchical Deep Architectureではベスト (acc:90%) | https://arxiv.org/pdf/1706.08574.pdf | ||||
(Github) | TLR | SSD with Mobilenet v1 | 38 ms, mAP:0.60 | Hierarchical Deep Architecture | https://github.com/bosch-ros-pkg/bstld/blob/master/tf_object_detection/configs/ssd_mobilenet_v1.config | ||||||
2019 | Yunnan University | Traffic Sign Recognition by Combining Global and Local Features Based on Semi-supervised Classification | TSR | CNN, SVM+Multiple feature fusion | P-R系の評価指標, CH+EF+HOGの融合 分類のみみたい | labeled dataの割合を振って評価 | https://www.researchgate.net/profile/Yun_Yang58/publication/336114360_Traffic_Sign_Recognition_by_Combining_Global_and_Local_Features_Based_on_Semi-supervised_Classification/links/5d8ecc56a6fdcc2554a10bf8/Traffic-Sign-Recognition-by-Combining-Global-and-Local-Features-Based-on-Semi-supervised-Classification.pdf | ||||
2019 | NVIDIA | REAL - TIME DETECTION OF LANES AND BOUNDARIES BY AUTONOMOUS VEHICLES | Lane Detection | 特許、CNN、binary/multi-classセグメンテーション+曲線フィッティング、ROI分割とDown Samplingの前処理組合せ | https://patentimages.storage.googleapis.com/48/10/c2/0598aac7501408/US20190266418A1.pdf | ||||||
2019 | NATIONAL CHIAO TUNG UNIVERSITY | LANE LINE DETECTION METHOD | Lane Detection | 特許、輝度から天候判断して適応的閾値2値化 | https://patentimages.storage.googleapis.com/15/80/2a/b2bdca577a1afb/US20190279003A1.pdf | ||||||
2018 | PLA Army Engineering University, | Multi-Object Detection in Traffic Scenes Based on Improved SSD | Object Detection | SSD改良 | |||||||
2019 | Toward Convolutional Blind Denoising of Real Photographs | NR | CBDNet | http://openaccess.thecvf.com/content_CVPR_2019/papers/Guo_Toward_Convolutional_Blind_Denoising_of_Real_Photographs_CVPR_2019_paper.pdf | |||||||
2019 | Megvii | Learning Raw Image Denoising with Bayer Pattern Unification and Bayer Preserving Augmentation | NR | https://arxiv.org/pdf/1904.12945.pdf | https://github.com/Jiaming-Liu/BayerUnifyAug. | ||||||
2019 | UIUC 2 Oxford 3 Megvii (Face++) 4Stony Brook 5 IBM | When AWGN-based Denoiser Meets Real Noises | NR | real noiseをブラックボックスとして扱う | https://arxiv.org/pdf/1904.03485.pdf | ||||||
2007 | Tampere University of Technology | Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data | NR | 輝度依存ノイズモデルの先駆 | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.164.1943&rep=rep1&type=pdf | ||||||
2019 | Overview and empirical analysis of ISP parameter tuning for visual perception in Autonomous Driving | HOG+LBP+LSVM, IOU | ISPの認識への影響という点では参考になる と思いきや、CLAHEで特徴点マッチングが減るとかどうでもいい内容 | https://www.researchgate.net/publication/334537057_Overview_and_empirical_analysis_of_ISP_parameter_tuning_for_visual_perception_in_Autonomous_Driving | |||||||
2019 | University of Michigan | Shadow Transfer: Single Image Relighting For Urban Road Scenes | Relighting | cast shadowもrelightingできるが性能がへぼい | https://arxiv.org/pdf/1909.10363.pdf | ||||||
2019 | Tsinghua University | Traffic Sign Detection based on SSD | TSR | SSD | 有料 | precision and recall on the test data set are 91.09%, and 88.06% | https://dl.acm.org/citation.cfm?id=3351988 | ||||
2017 | Tribhuvan University | A Novel Approach to Traffic Sign Localization and Recognition Based on Single Shot Multibox Detector | TSR | SSD (ZFNet) | mAP:0.81 | MSERs | https://pdfs.semanticscholar.org/590a/bd947003d26f547c022182f7dca13b0353af.pdf | ||||
2018 | Qualcomm | Simultaneous Traffic Sign Detection and Boundary Estimation using Convolutional Neural Network | TSR | Vertex regression, Faster RCNN and SSD | mAP:75.4, 4.3FPS, 0.39GMAC(SqueezeNet) | https://arxiv.org/pdf/1802.10019.pdf | |||||
2019 | University of Freiburg, Bosch | UNDERSTANDING AND ROBUSTIFYING DIFFERENTIABLE ARCHITECTURE SEARCH | NAS | DARTS ADA, premature convergence in the weights’ space | DARTS | https://arxiv.org/pdf/1909.09656.pdf | |||||
2019 | University of Science and Technology of China, JD AI Research | Scheduled Differentiable Architecture Search for Visual Recognition | NAS | Visual Recognition, Schedule DAS (SDAS), v | 0.25GPUdays | SENet-3D | https://arxiv.org/pdf/1909.10236.pdf | ||||
2019 | University of Science and Technology of China, Microsoft Research Asia | Understanding and Improving One-shot Neural Architecture Optimization | NAS | NAO-V2 | 1GPUdays | NAO-WS, ProxylessNAS | https://arxiv.org/pdf/1909.10815.pdf | ||||
2019 | Karlsruhe Institute of Technology (KIT) | FAIRnets Search - A Prototype Search Service to Find Neural Networks | NAS | RDF and OWL | http://ceur-ws.org/Vol-2451/paper-19.pdf | http://km.aifb.kit.edu/services/fairnets/. | |||||
2019 | Megvii, Tsinghua University, Xi’an Jiaotong University | Double Anchor R-CNN for Human Detection in a Crowd | Object Detection | Joint NMS(頭部・胴体分ける), human in a crowd | miss rates (MR) of 51.79pp | https://arxiv.org/pdf/1909.09998.pdf | |||||
2019 | University of Chinese Academy of Sciences | A System-Level Solution for Low-Power Object Detection | Object Detection | MobileNet | mAP of 66.4 verified on the PASCAL VOC 2012, 202.76GOPs(Xilinx Zynq UltraScale+ MPSoC) | https://arxiv.org/pdf/1909.10964.pdf | |||||
2019 | Improving CNN-based Planar Object Detection with Geometric Prior Knowledge | Object Detection | YOLO or MASK RCNN | 斜めの平面物体 D情報でRGBを平行化 | https://arxiv.org/pdf/1909.10245.pdf | ||||||
2019 | City University of Hong Kong | CNN-based RGB-D Salient Object Detection: Learn, Select and Fuse | Object Detection | Hierarchical Cross-modal Distillation | RGBをTeacherに、DをStudentに入力。 各層でFusion | https://arxiv.org/pdf/1909.09309.pdf | |||||
2019 | University of Illinois at Urbana-Champaign, Inspirit IoT, Inc, IBM Research | SKYNET: A HARDWARE-EFFICIENT METHOD FOR OBJECT DETECTION AND TRACKING ON EMBEDDED SYSTEMS | Object Detection, Model Compression | Har56th IEEE/ACM Design Automation Conference (DAC-SDC), dware-Aware Neural Network Search | NAS | 67.33 frames per second (FPS) on a TX2 | https://arxiv.org/pdf/1909.09709.pdf | ||||
2019 | University of Maryland College Park, CCDC Army Research Laboratory | WASSERSTEIN DISTANCE BASED DOMAIN ADAPTATION FOR OBJECT DETECTION | Object Detection | unsupervised domain adaptation | 霧、Global Features/gradient reversal layer (GRL)に対しDiscriminator | https://arxiv.org/pdf/1909.08675.pdf | |||||
2013 | 車載カメラを用いた自転車検出システムの研究 | HOGとCoHOGの速度比較、さらに倍速 | https://core.ac.uk/download/pdf/147425919.pdf | ||||||||
2014 | The University of Tokyo | Extended Co-occurrence HOG with Dense Trajectories for Fine-grained Activity Recognition | ECoHOG | http://www.hirokatsukataoka.net/pdf/sice14_kataoka_ecohoganalysis.pdf http://www.hirokatsukataoka.net/pdf/accv14_kataoka_finegrainedactivityrecognition.pdf | |||||||
2014 | Hanyang University | Outdoor Place Recognition in Urban Environments using Straight Lines | Canny, MSLD | OpenCVのcv::ximgproc::FastLineDetector | http://cvlab.hanyang.ac.kr/~jwlim/files/icra14linerec.pdf | ||||||
2013 | Ritsumeikan University | Data Transfer Matters for GPU Computing | hardware-assisted direct memory access | GPU高速化参考 | http://www.ertl.jp/~shinpei/papers/icpads13.pdf | ||||||
2017 | Kyushu Institute of Technology | Evaluation of Hardware Oriented MRCoHOG using Logic Simulation | Pedestrian | MRCoHOG | https://pdfs.semanticscholar.org/7501/3eb524b894ff965db53c1a5ad44523baf53f.pdf | ||||||
2019 | Google Research | EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks | NAS | Compound Model Scaling | 下記引用 推論は速くなったが学習時間は不明 depth,width,resolutionの積に制約 ACC*FLOPS^-0.07をobjective | 84.4% top-1 / 97.1% top-5 accuracy on ImageNet, 8.4x smaller and 6.1x faster | https://arxiv.org/pdf/1905.11946.pdf | https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet | |||
2019 | University of Tsukuba & RIKEN AIP,Yokohama National University | Adaptive Stochastic Natural Gradient Method for One-Shot Neural Architecture Search | NAS | ASNG-NAS | 日経ロボティクスで取り上げられていた | 0.11GPUdays | http://proceedings.mlr.press/v97/akimoto19a/akimoto19a.pdf | https://www.skillupai.com/blog/NNpaper/ | |||
2019 | Duke University | DASNet: Dynamic Activation Sparsity for Neural Network Efficiency Improvement | Model Compression | winners-take-all (WTA) dropout | +0.26%@15.9%MAC(ResNet-50) | https://arxiv.org/pdf/1909.06964.pdf | |||||
2019 | University of Toronto | Monocular 3d object detection leveraging accurate proposals and shape reconstruction | Object Detection | 下記引用 | MultiFusion | http://openaccess.thecvf.com/content_CVPR_2019/papers/Ku_Monocular_3D_Object_Detection_Leveraging_Accurate_Proposals_and_Shape_Reconstruction_CVPR_2019_paper.pdf | |||||
2019 | The University of Adelaide, Bytedance AI Lab | Task-Aware Monocular Depth Estimation for 3D Object Detection | Depth, Object Detection | Depth評価指標が参考になる(absRel, SILog etc.) | MonoPSR | https://arxiv.org/pdf/1909.07701.pdf | |||||
2019 | Alibaba | Line as object: datasets and framework for semantic line segment detection | Lane Detection | hourglass104 0. | MS COCO mAPでKAIST SLS datasetsを評価 評価方法が詳しく参考になる IoUの代替としてACL(8) | mAP:0.7353, 36fps( GTX 1080Ti) | https://arxiv.org/pdf/1909.06591.pdf | ||||
2019 | Skoltech, Samsung | Recognition of Russian traffic signs in winter conditions. Solutions of the “Ice Vision” competition winners | TSR | Cascade R-CNN | 悪天候下のコンペ 標識の空間分布、やはり上半分にしかない 標識は反転すると正しくないので、finetuneではflip DAをOFFる | https://arxiv.org/pdf/1909.07311.pdf | https://github.com/icevision/solutions | ||||
2019 | Technion, Intel | Thanks for Nothing: Predicting Zero-Valued Activations with Lightweight Convolutional Neural Networks | Model Compression | zero activation predictor (ZAP) | MAC reduction:23.1%~ | https://arxiv.org/pdf/1909.07636.pdf | |||||
2019 | Chubu University | Attention Branch Network: Learning of Attention Mechanism for Visual Explanation | Attention branch | 下記引用,直観に整合したattention mapが得られている | Average of accuracy:91.07 | https://arxiv.org/pdf/1812.10025.pdf | |||||
2019 | TAIST, 東工大 | New Perspective of Interpretability of Deep Neural Networks | Interpretability | human predictability(easiness for the human to predict the change of the inference when perturbing the model), Human-in-the-loop (HITL), Channel Pruning | Average of accuracy:90.72 | Attention Branch Network (ABN) | https://arxiv.org/pdf/1909.07156.pdf | ||||
2019 | Huawei,Tsinghua University | DARTS+: Improved Differentiable Architecture Search with Early Stopping | NAS | DRATS, early stopping paradigm | DARTSのskip-connectが多発する問題(Collapse Issue)を解決 | 0.2GPUdays(DARTSのx20) 2.32% test error on CIFAR10, 14.87% on CIFAR100, and 23.7% on ImageNet | XNAS | https://arxiv.org/pdf/1909.06035.pdf | |||
2019 | JD AI Research | Customizable Architecture Search for Semantic Segmentation | NAS, Semantic Segmentation | Multi-Scale Cell Search | 下記引用,速度は最高、Normal/Reduction直列、パラメータ数制御 | http://openaccess.thecvf.com/content_CVPR_2019/papers/Zhang_Customizable_Architecture_Search_for_Semantic_Segmentation_CVPR_2019_paper.pdf | |||||
2019 | SenseTime Research, Zhejiang University | Graph-guided Architecture Search for Real-time Semantic Segmentation | NAS, Semantic Segmentation | graph convolution network (GCN) | mIOU:71.9, 142fps(CamVid, 720x960) | CAS | https://arxiv.org/pdf/1909.06793.pdf | ||||
2019 | MOBILEYE | SYSTEMS AND METHODS FOR CURB DETECTION AND PEDESTRIAN HAZARD ASSESSMENT | 縁石認識 warp functionによる段差推定 高さ既知? | https://patentimages.storage.googleapis.com/d1/f1/93/b1d16dd174067f/US20190251375A1.pdf | |||||||
2019 | TOYOTA | TRAFFIC SIGNAL RECOGNITION DEVICE | TLR | mapを使った信号機認識の特許 | http://scholar.google.co.jp/scholar_url?url=https://patentimages.storage.googleapis.com/00/53/57/0dce5a8ae16c91/US20190244041A1.pdf&hl=ja&sa=X&d=12831132657367409302&scisig=AAGBfm1pbt2GeUesHehWbL4nIarnsJkHKA&nossl=1&oi=scholaralrt&hist=5RIGpeIAAAAJ:13950008414835878474:AAGBfm2SZXPONllfi4fVPoFhzTloxU35Gg | ||||||
2019 | The Hong Kong University of Science and Technology | Multi-Sensor 3D Object Box Refinement for Autonomous Driving | Object Detection | 3D, Frustum PC, stereo/3D box alignment | LiDAR/StereoとのFusion | Pseudo-3D | https://arxiv.org/pdf/1909.04942.pdf | ||||
2019 | Norwegian University of Science and Technology | Automated Detecting and Placing Road Objects from Street-level Images | Object Detection | attributed topological binary tree (ATBT) | PSPNetがDeepLabv3+を、YOLOがFasterを上回る | https://arxiv.org/ftp/arxiv/papers/1909/1909.05621.pdf | |||||
2019 | Peking University, Huawei | CARS: Continuous Evolution for Efficient Neural Architecture Search | NAS | evolution, Parameter Warmup, pNSGA-III | 0.4GPU days, AmoebaNEt並み精度 | https://arxiv.org/pdf/1909.04977.pdf | |||||
2016 | Robust Uncalibrated Stereo Rectification with Constrained Geometric Distortions (USR-CGD) | Rectification | 異焦点, 変換後の4つの矩形の幾何的定量指標をlossに加えた最適化(non-convex) カメラ中心の移動を許容する平行化により画像のゆがみを抑えつつ異焦点に対応する | https://arxiv.org/pdf/1603.09462.pdf | http://mcl.usc.edu/mcl-ss-database/ | ||||||
2017 | Karlsruhe Institute of Technology | Online Stereo Camera Calibration From Scratch | Camera Calibration | Motion Estimation | あいにくgithubは空っぽ、同焦点距離 再投影誤差ベースで動き推定 アルゴ詳細が書いてない | https://www.mrt.kit.edu/z/publ/download/2017/rehder_iv17.pdf | https://github.com/KIT-MRT. | ||||
2007 | U n iv e rs ity o f N o rth C aro lin a | Real-time plane-sweeping stereo with multiple sweeping directions | Depth | plane-sweeping, GPU | 20fps | https://inf.ethz.ch/personal/pomarc/pubs/GallupCVPR07.pdf | |||||
2015 | Dalian University of Technology | Improved camera calibration method based on perpendicularity compensation for binocular stereo vision measurement system | Camera Calibration | accuracy:99.91% | https://www.osapublishing.org/DirectPDFAccess/40850022-B308-C206-F9337F0A1F6DD36B_319955/oe-23-12-15205.pdf?da=1&id=319955&seq=0&mobile=no | ||||||
2014 | ETH | Real-Time Direct Dense Matching on Fisheye Images Using Plane-Sweeping Stereo | Stereo | ZNCC, sub-pixel interpolation | Plane-Sweepingによる魚眼のステレオ 平行化でない理由は、マルチカメラへの拡張、GPUによる実時間処理 | https://www.researchgate.net/profile/Lionel_Heng/publication/282927545_Real-Time_Direct_Dense_Matching_on_Fisheye_Images_Using_Plane-Sweeping_Stereo/links/5628be5b08ae22b1702ed18e.pdf | |||||
1999 | Microsoft Research | Computing Rectifying Homographies for Stereo Vision | Camera Calibration | Epipolar Geometry, Rectification, Shearing Transform | fundamental matrixをHomographyの外積に、Homography行列をprojective/shearing/similarityの積に分解 Homographyは平行化画像の座標に変換する行列 | http://dev.ipol.im/~morel/Dossier_MVA_2011_Cours_Transparents_Documents/2011_Cours7_Document2_Loop-Zhang-CVPR1999.pdf | |||||
2002 | Middlebury College, Microsoft Research | A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms | Stereo | http://vision.middlebury.edu/stereo/taxonomy-IJCV.pdf | |||||||
2012 | Raytrix GmbH | Single Lens 3D-Camera with Extended Depth-of-Field | Depth | plenopticの理論が詳しい | https://www.researchgate.net/profile/Christian_Perwass/publication/258713151_Single_Lens_3D-Camera_with_Extended_Depth-of-Field/links/575e978b08aec91374b3d90a.pdf | ||||||
2019 | Peking University, Johns Hopkins University | TDAPNet: Prototype Network with Recurrent Top-Down Attention for Robust Object Classification under Partial Occlusion | Image Classification | attention mapを作って上流にフィードバック | https://arxiv.org/pdf/1909.03879.pdf | ||||||
1998 | University of Parma | GOLD: A Parallel Real-Time Stereo Vision System for Generic Obstacle and Lane Detection | Lane Detection | STEREO INVERSE PERSPECTIVE MAPPING | (u,v)⇔(x,y)の解析的表現 輝度画像の1画素飛ばしの微分 StereoのPolar histogramsの波形で障害検知 | http://www.ce.unipr.it/people/bertozzi/publications/cr/ieee.ip.gold.pdf | |||||
2019 | NVIDIA | VACL: Variance-Aware Cross-Layer Regularization for Pruning Deep Residual Networks | Model Compression | ~79.5%prune | https://arxiv.org/pdf/1909.04485.pdf | ||||||
2019 | University of Oxford | Geometry-Aware Video Object Detection for Static Cameras | Object Detection | Geometry Inputからattentionを作る | https://arxiv.org/pdf/1909.03140.pdf | ||||||
2010 | Distance determination for an automobile environment using Inverse Perspective Mapping in OpenCV | Depth | IPM | 下記からの引用 実質簡易単眼測距(y座標)と同じ | https://www.researchgate.net/publication/224195999_Distance_determination_for_an_automobile_environment_using_Inverse_Perspective_Mapping_in_OpenCV | ||||||
2019 | New York University | Learning Object-Specific Distance From a Monocular Image | Depth | Distance/Keypoint Regressor | 単眼測距,KITTIの距離頻度参考 | 16.2ms, RMSE:6.87m(KITTI), 10.511(nuScenes) | https://arxiv.org/pdf/1909.04182.pdf | ||||
2019 | Traffic Sign Detection and Classification around the World | TSR | 世界の標識を集めたデータセット構築 | https://arxiv.org/pdf/1909.04422.pdf | https://www.mapillary.com/dataset/trafficsign | ||||||
2017 | Daimler, University of Freiburg | Sparsity Invariant CNNs | depth completion | sparse inputs | 全層にマスクの積を含むSparse Convolutionを行う LiDARのsparseな点群から密なデプスマップを作るのが目的 | http://www.cvlibs.net/publications/Uhrig2017THREEDV.pdf | |||||
2017 | Facebook AI Research, University of Oxford | 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks | Semantic Segmentation | Sparse Convolution operate on spatially-sparse input data | スパースな入力に対応 | https://arxiv.org/pdf/1711.10275.pdf | https://github.com/facebookresearch/SparseConvNet | ||||
2019 | Stanford | 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks | Object Detection | Point Cloud | http://openaccess.thecvf.com/content_CVPR_2019/papers/Choy_4D_Spatio-Temporal_ConvNets_Minkowski_Convolutional_Neural_Networks_CVPR_2019_paper.pdf | https://github.com/StanfordVL/MinkowskiEngine | |||||
2019 | Google Brain | Weight Agnostic Neural Networks | NAS | topology search algorithm (NEAT), multi-objective optimization. | パラメータはランダムなまま(学習しない)で, ネットワーク構造(接続)を進化させる | https://arxiv.org/pdf/1906.04358.pdf | https://weightagnostic.github.io/ | https://ai.googleblog.com/2019/08/exploring-weight-agnostic-neural.html http://cympfh.cc/paper/weightanogstic.html https://www.slideshare.net/DeepLearningJP2016/dlweight-agnostic-neural-networks | |||
2019 | Facebook AI Research | Exploring Randomly Wired Neural Networks for Image Recognition | NAS | 一定ルールの中では乱数で設計した方が高性能という結果 | https://arxiv.org/pdf/1904.01569.pdf | ||||||
2019 | Xiaomi | MoGA: Searching Beyond MobileNetV3 | NAS | MobileNetV3を速度・精度両面で上回る | https://arxiv.org/pdf/1908.01314.pdf | https://github.com/xiaomi-automl/MoGA | |||||
2012 | University of Patras | Random-Walker Monocular Road Detection in Adverse Conditions Using Automated Spatiotemporal Seed Selection | Free Space Detection | c1c2c3 color space, Random Walker Algorithm (RWA) | shadowに強いOFのための色変換 | https://www.researchgate.net/publication/233752622_Random-Walker_Monocular_Road_Detection_in_Adverse_Conditions_Using_Automated_Spatiotemporal_Seed_Selection | |||||
2018 | VNU University of Engineering and Technology | Real-time Lane Marker Detection Using Template Matching with RGB-D Camera | Lane Detection | half-binary format, template matching | 50ms(Jetson TX2) 独自データでTP53% | https://arxiv.org/pdf/1806.01621.pdf | |||||
2018 | University of Bristol | Real-Time Stereo Vision-Based Lane Detection System | Lane Detection | RANSAC, LSF, V-disparity map, dynamic programming road surface detection | https://arxiv.org/pdf/1807.02752.pdf | ||||||
2018 | Lane Detection Based on Inverse Perspective Transformation and Kalman Filter | Lane Detection | IPM, Kalman Filter | DoG?前処理 IPMは台形か | http://itiis.org/digital-library/manuscript/file/1921/TIIS+Vol+12,+No+2-6.pdf | ||||||
2019 | KU Leuven | End-to-end Lane Detection through Differentiable Least-Squares F | Lane Detection | ERFNet | 高速の評価 オクルージョンに強い | https://arxiv.org/pdf/1902.00293.pdf | https://github.com/wvangansbeke/LaneDetection_End2End | ||||
2019 | University of Chinese Academy of Sciences, SenseTime | Efficient Neural Architecture Transformation Search in Channel-Level for Object Detection | NAS | neural architecture transformation search(NATS) Channel-wise Selection & Concat | Object Detectionで最多・最高解像度の学習データ(COCO, 118k, 800x1200) 20GPU days | +0.5% | https://arxiv.org/pdf/1909.02293.pdf | ||||
2019 | University of Chinese Academy of Sciences, SenseTime | POD: Practical Object Detection with Scale-Sensitive Network | Object Detection | scale variance | 100ms, 38.5%@COCO AP | Cascade R-CNN, CornerNet | https://arxiv.org/pdf/1909.02225.pdf | ||||
2019 | Chinese Academy of Sciences | Rethinking the Number of Channels for Convolutional Neural Networks | NAS | Widening a Convolutional Layer, Evolutionary Algorithm | +0.5% on CIFAR-10 and a∼2.33% on CIFAR-100 | https://arxiv.org/pdf/1909.01861.pdf | |||||
2019 | University of California | What Happens on the Edge, Stays on the Edge: Toward Compressive Deep Learning | Compressive sensing | エッジでやる動機は、privacy,security,通信コスト ROIを重点的にサンプリング | https://arxiv.org/pdf/1909.01539.pdf | ||||||
2010 | Cranfield University | Automatic real-time road marking recognition using a feature driven approach | Road Marking Recognition | IPM, complexity rejection, orientation normalisation, Shape isolation(backtracking approach [17]) | Canny+Houghで消失点推定 VPの位置からyaw,pitch推定→point to point IPM MLPで分類 orientation normalisationをPython実装した | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.227.8116&rep=rep1&type=pdf | |||||
2011 | University of Chile | Harris 3D: a robust extension of the Harris operator for interest point detection on 3D meshes | Harris 3D | http://personales.dcc.uchile.cl/~isipiran/papers/SB11b.pdf | |||||||
2006 | YAMAHA | MONOCULAR-MOTION-STEREO DISTANCE ESTIMATION METHOD, AND MONOCULAR-MOTION-STEREO DISTANCE ESTIMATION APPARATUS | Motion Stereo | 特許 (10)式:OFとカメラの並進量から推定 | https://patentimages.storage.googleapis.com/45/4d/48/72017c52acecc4/EP3009789A1.pdf | ||||||
2019 | Universitas Prima Indonesia | Traffic sign detection using histogram of oriented gradients and max margin object detection | TSR | HOG, SVM, Max Margin Object Detection (MMOD) | https://iopscience.iop.org/article/10.1088/1742-6596/1230/1/012098/pdf | ||||||
2004 | Implementation of Inverse Perspective Mapping Algorithm for the Development of an Automatic Lane Tracking System | IPM | 下記引用 | 有料 | |||||||
2016 | Dongguk University | Recognition of Damaged Arrow-Road Markings by Visible Light Camera Sensor Based on Convolutional Neural Network | Road Marking Recognition | CNN | |||||||
2007 | Inria | Deeper understanding of the homography decomposition for vision-based control | Camera Calibration | homography decomposition | https://hal.inria.fr/inria-00174036/PDF/RR-6303.pdf | ||||||
2016 | University of Applied Sciences Upper Austria | Zhang’s Camera Calibration Algorithm: In-Depth Tutorial and Implementation | Camera Calibration | Zhangの方法の完全な理解 | http://staff.fh-hagenberg.at/burger/publications/reports/2016Calibration/Burger-CameraCalibration-20160516.pdf | ||||||
2005 | International Institute of Information Technology | A Survey of Planar Homography Estimation Techniques | Camera Calibration | two views, Planar homography, epipolar geometry, DLT | H=K’Rinv(K) 点・線・円弧のマッチング | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.102.321&rep=rep1&type=pdf | |||||
2018 | A novel solution in the simultaneous deep optimization of RGB-D camera calibration parameters using metaheuristic algorithms | Camera Calibration | Kinect v1,v2 | https://dergipark.org.tr/tr/download/article-file/728926 | |||||||
2001 | Just Research | Smarter Presentations: Exploiting Homography in Camera-Projector Systems | Camera Calibration | プロジェクタのdist補正 | https://www.cs.cmu.edu/~rahuls/pub/iccv2001-rahuls.pdf | ||||||
2019 | Michigan State University, AInnovation | HM-NAS: Efficient Neural Architecture Search via Hierarchical Masking | NAS | gradient-based | 0.45~1.4GPU days | DARTS | https://arxiv.org/pdf/1909.00122.pdf | ||||
2019 | Huawei | MANAS: Multi-Agent Neural Architecture Search | NAS | multi-agent problem memory requirements (1/8th of state-of-the-art | 0.8~4GPU days | DARTS | https://arxiv.org/pdf/1909.01051.pdf | ||||
2019 | ETH | EBPC: Extended Bit-Plane Compression for Deep Neural Network Inference and Training Accelerators | Acceleration | TABLE I. COMPARISON OF FEATURE MAP SIZES AND PARAMETER COUNT OF MODERN DNNS が有用 | average compression ratio of 4× on VGG-16 (+67%), 2.2× on MobileNetV2 (+30%) | https://arxiv.org/pdf/1908.11645.pdf | |||||
2019 | Valeo | FisheyeMODNet: Moving Object detection on Surround-view Cameras for Autonomous Driving | Object Detection | fisheye surround-view s | 2フレーム入力して動体マスクを出力 | 15fps(1 Megapixel images) | https://arxiv.org/pdf/1908.11789.pdf | ||||
2019 | Nanjing University of Posts and Telecommunications | Feature Fusion and Adversary Occlusion Networks for Object Detection | Object Detection | FFAN Adversary Occlusion Network | small objects対策 FasterがDSしていくのに対し、後段のdeconvを付加しUNet的構造にする 効果が微妙 | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8821288 | |||||
2019 | Middle East Technical University (METU) | Imbalance Problems in Object Detection: A Review | Object Detection | Scale/Objective/Class/Spatial imbalance | Imbalanceに特化したサーベイ | https://arxiv.org/pdf/1909.00169.pdf | |||||
2019 | Object detection from a few LIDAR scanning planes | Object Detection | Fourier descriptor | 3層CNN+FC KITTI dataset; B - Budapest dataset [7] | Precision:75% | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8818349 | |||||
2005 | Xi'an Jiaotong University | Comparison and choice of models in tracking target with coordinated turn motion | Tracking | UKF | 下記で出てきたAugmented Coordinated Turn ACTはKFの遷移行列を(4)で与える | https://www.researchgate.net/publication/4221183_Comparison_and_choice_of_models_in_tracking_target_with_coordinated_turn_motion | |||||
2019 | University of the Bundeswehr Munich | Combining Deep Learning and Model-Based Methods for Robust Real-Time Semantic Landmark Detection | Object Detection | LiDAR, Segmented Point Clouds, Greedy Dirichlet Process Filter (GDPF) of [16], the Augmented Coordinated Turn [38] | Landmarkということで注目 しかし対象はTree,Bush RGB + LiDAR | 92ms(GTX 1060) | ComplexYOLO [32] and Voxelnet | https://arxiv.org/pdf/1909.00733.pdf | |||
2019 | Jiangsu University | Construction and Simulation of Rear-end Conflicts Recognition Model Based on Improved TTC Algorithm | TTC | time to collision in the work zone (WTTC) | WTTCの提案 | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8818033 | |||||
2008 | SANYO | Precise Top View Image Generation | IPM | リアカメラのキャリブの簡易化? | http://www.cc.kyoto-su.ac.jp/~kano/pdf/paper/2008%20IEICE%20Top%20View.pdf | ||||||
2018 | Shanghai University of Engineering Science | Lane Detection Based on Connection of Various Feature Extraction Methods | Lane Detection | Canny, EKF, polar Hough | http://downloads.hindawi.com/journals/am/2018/8320207.pdf | ||||||
2019 | Shandong University of Science and Technology, | Research on Lane a Compensation Method Based on Multi-Sensor Fusion | Lane Detection | GPS,CAN,IMUとのFusion ラインを3次で表現 | https://www.researchgate.net/publication/332177262_Research_on_Lane_a_Compensation_Method_Based_on_Multi-Sensor_Fusion | ||||||
2014 | University of California San Diego | On Performance Evaluation Metrics for Lane Estimation | Lane Detection | ライン定量評価のLPDの出典 ELASで適用 | http://cvrr.ucsd.edu/publications/2014/SatzodaTrivedi_ICPR2014.pdf | ||||||
2019 | The Chinese University of Hong Kong, SenseTime | Learning Lightweight Lane Detection CNNs by Self Attention Distillation | Lane Detection | Activation-based attention distillation | LDのSOTA インターンで追試 | ENet, ERFNet, SCNN | https://arxiv.org/pdf/1908.00821v1.pdf | https://github.com/cardwing/Codes-for-Lane-Detection | https://paperswithcode.com/paper/learning-lightweight-lane-detection-cnns-by | ||
2019 | Chongqing University | FFBNET : LIGHTWEIGHT BACKBONE FOR OBJECT DETECTION BASED FEATURE FUSION BLOCK | Object Detection | SSD+VGGとコンパラ | 73.54 mAP at speed of 185 FPS(Pascal VOC) | MobileNet-SSD | https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8803683 | https://github.com/fanbinqi/FFBNet. | |||
2017 | Rice University | Learning to Invert: Signal Recovery via Deep Convolutional Networks | Compressed sensing | DLで復元 従来より数百倍高速 画像を見るとDAMPの方が全然きれい 圧縮率はより上げられる 入力をFC(ΦTをかける操作に相当)で画像の次元に戻すのがポイント 3 layer CNN, 11x11 kernel 先行例[11]に倣い画像をブロック分割 | DAMP | https://arxiv.org/pdf/1701.03891.pdf | |||||
2017 | Nvidia Research, Simon Fraser University | Polarimetric Multi-View Stereo | Polarimetric imaging | 3種類の反射:(un)polarized diffuse reflection, polarized specular reflection specularは位相がπ/2ずれる MVSと組み合わせ形状推定 重要な発見として、無偏光照明に対しては、specular/diffuseの比率によらず、偏光角に対し正弦波となる | https://research.nvidia.com/sites/default/files/pubs/2017-07_Polarimetric-Multi-view-Stereo/PolarStereo.pdf | ||||||
2012 | USC Institute for Creative Technologies | Estimating Surface Normals from Spherical Stokes Reflectance Fields | Shape from polarisation | 下記からの引用 | http://ict.usc.edu/pubs/Estimating%20Surface%20Normals%20from%20Spherical%20Stokes%20Reflectance%20Fields.pdf | ||||||
2017 | Polarization imaging reflectometry in the wild | Polarimetric imaging | Transmitted Radiance Sinusoid (TRS), Reflectance extraction, per-pixel index of refraction | 0◦, 45◦and 90◦のデータにSVDで正弦波フィッティング specular refectance推定で不定性解消のためカラーチャートの黒縁を使う | http://www.ilya.o-x-t.com/publications/polarization-reflectometry-accepted-small.pdf | ||||||
2019 | Depth from a polarisation + RGB stereo pair | Stereo | ステレオの一方を偏光にすることで微細凹凸を取得 | http://openaccess.thecvf.com/content_CVPR_2019/papers/Zhu_Depth_From_a_Polarisation__RGB_Stereo_Pair_CVPR_2019_paper.pdf | |||||||
2018 | University of York | Height-from-Polarisation with Unknown Lighting or Albedo | Shape from polarisation | Degree of polarisation | orthographic(正投影)の制約 | https://www-users.cs.york.ac.uk/wsmith/papers/SmRaTo18_final.pdf | |||||
1982 | Correlation Between Albedo and Polarization Characteristics of the Moon - Application of Digital Image Processing | http://adsabs.harvard.edu/full/1982SvA....26..345K | |||||||||
2019 | Wuhan University | Infrared and visible image fusion methods and applications: A survey | Image fusion | ||||||||
2015 | KAIST | Multispectral Pedestrian Detection: Benchmark Dataset and Baseline | Object Detection | 人は長波長9.3µmを放射 FLIR-A35(Spectral Range 7.5 – 13 µm) | http://openaccess.thecvf.com/content_cvpr_2015/papers/Hwang_Multispectral_Pedestrian_Detection_2015_CVPR_paper.pdf | http://rcv.kaist.ac.kr/multispectral-pedestrian | |||||
2019 | Megvii Technology, Tsinghua University | Learning Raw Image Denoising with Bayer Pattern Unification and Bayer Preserving Augmentation | Noise reduction | https://arxiv.org/pdf/1904.12945.pdf | https://github.com/Jiaming-Liu/BayerUnifyAug | ||||||
2015 | University of Toronto, Qualcomm | High Resolution Photography with an RGB-Infrared Camera | Demosaicing | RGB-IRのCFAで画像復元 pixel multiplexing, channel crosstalk and chromatic aberrationsの3つの課題に対処 | https://www.cs.toronto.edu/~kyros/pubs/15.iccp.rgbi.pdf | ||||||
2017 | The Chinese University of Hong Kong, SenseTime | Spatial As Deep: Spatial CNN for Traffic Scene Understanding | Lane Detection | SCNN | 中間層でH/V方向に分割するのが特徴 白線がなくても線を引く | 133ms, F1:71 | https://arxiv.org/pdf/1712.06080.pdf | ||||
2019 | Agnostic Lane Detection | Lane Detection | ENet | 13.4ms, F1:68.8 | https://arxiv.org/pdf/1905.03704.pdf | ||||||
2018 | General Motors | 3D-LaneNet: end-to-end 3D multiple lane detection | Lane Detection | intra-network inverse-perspective mapping, anchor-based lane representation | 勾配も同時推定 | Error near(0-30m) (cm):10.8, Error far(30-80m) (cm):30 (tuSimple) | https://arxiv.org/pdf/1811.10203v1.pdf#page=1 | ||||
2018 | Cranfield University | Advances in Vision-Based Lane Detection: Algorithms, Integration, Assessment, and Perspectives on ACP-Based Parallel Vision | Lane Detection | LDのサーベイとして有用 | http://www.ieee-jas.org/fileZDHXBEN/journal/article/zdhxbywb/2018/3/PDF/jas-5-3-645.pdf | ||||||
Toyota | Image Processing Technology for Rear View Camera (1) : Development of Lane Detection System | extended Hough | Houghの高速化の参考、LUT重みづけ | https://www.tytlabs.com/english/review/rev382epdf/e382_031takahashi.pdf | |||||||
2019 | Research on Lane a Compensation Method Based on Multi-Sensor Fusion | Lane Detection | |||||||||
2019 | Tsinghua University, SenseTime, | DrivingStereo: A Large-Scale Dataset for Stereo Matching in Autonomous Driving Scenarios | Depth | over 180k images, KITTIの100倍以上 | http://openaccess.thecvf.com/content_CVPR_2019/papers/Yang_DrivingStereo_A_Large-Scale_Dataset_for_Stereo_Matching_in_Autonomous_Driving_CVPR_2019_paper.pdf | https://drivingstereo-dataset.github.io | |||||
2019 | Self-Attention Generative Adversarial Networks | GAN | Self-Attention Generative Adversarial Network (SAGAN), long-range dependency modeling, two timescale update rule (TTUR), spectral normalization | 下記からの引用 | Frechet Inception distance from 27.62 to 18.65 | https://arxiv.org/pdf/1805.08318.pdf | https://github.com/brain-research/self-attention-gan | ||||
2019 | Tel Aviv University | Single Image Depth Estimation Trained via Depth from Defocus Cues | Depth | Depth From Defocus, Circle-Of-Confusion, PSF Convolutional layer | http://openaccess.thecvf.com/content_CVPR_2019/papers/Gur_Single_Image_Depth_Estimation_Trained_via_Depth_From_Defocus_Cues_CVPR_2019_paper.pdf | ||||||
2016 | Kwangwoon University | A parallel camera image signal processor for SIMD architecture | ISP | ISP | 現像ISPへのSIMD応用 | https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-016-0137-2 | |||||
2018 | EAVISE, KU Leuven | GPU Accelerated ACF Detector | Object Detection | YOLOと同速度で低精度 | ~x3.8 | https://pdfs.semanticscholar.org/4ec4/392246a7760d189cd6ea48a81664cd2fe4bf.pdf | |||||
2019 | KAIST | Camera Exposure Control for Robust Robot Vision with Noise-Aware Image Quality Assessment | Auto-Exposure | Nelder-Mead (NM) method | ノイズを推定するので実質NR gradient,entropy,ノイズレベルの和であるIQMを最小化する 撮ってパラメータ調整してを反復するみたい | https://arxiv.org/pdf/1907.12646.pdf | |||||
2018 | University of Michigan, Baidu | Modeling Camera Effects to Improve Visual Learning from Synthetic Data | Data Augmentation | CGを学習に使うためのカメラ劣化モデル Virtual KITTI/GTAで評価 | http://openaccess.thecvf.com/content_ECCVW_2018/papers/11129/Carlson_Modeling_Camera_Effects_to_Improve_Visual_Learning_from_Synthetic_Data_ECCVW_2018_paper.pdf | ||||||
2018 | KAIST | Fast and Efficient Image Quality Enhancement via Desubpixel Convolutional Neural Networks | Super Resolution | FEQE, Perceptual index | DPED | http://openaccess.thecvf.com/content_ECCVW_2018/papers/11133/Vu_Fast_and_Efficient_Image_Quality_Enhancement_via_Desubpixel_Convolutional_Neural_ECCVW_2018_paper.pdf | https://github.com/thangvubk/FEQE.git | ||||
2016 | Arizona State University | Understanding How Image Quality Affects Deep Neural Networks | Image classification | 劣化のバリエーションが多様 | https://arxiv.org/pdf/1604.04004.pdf | ||||||
2018 | Fujitsu Laboratories Ltd.,Johns Hopkins University | Pushing the Limits of Unconstrained Face Detection: a Challenge Dataset and Baseline Results | Face Recognition | 劣化のバリエーションが多様 P-R評価のみ | https://arxiv.org/pdf/1804.10275.pdf | https://ufdd.info/ | |||||
2016 | TUBITAK BILGEM | How Image Degradations Affect Deep CNN-based Face Recognition? | Face Recognition | GoogLeNet | 劣化のバリエーションが多様 ノイズ・ブラーの敏感度が高い | https://arxiv.org/pdf/1608.05246.pdf | |||||
2017 | University of Chicago, TTI-Chicago | Examining the Impact of Blur on Recognition by Convolutional Networks | Semantic Segmentation | fine-tuned with a mix of sharp and range of defocus kernels (FT-Mix) Hamming distances network scale | blurの影響の検証 ただのdeblurより精度向上 Sharp & Camera Shakeによるfinetuneがベスト | https://arxiv.org/pdf/1611.05760.pdf | |||||
2017 | AGH University of Science and Technology | Image Recognition with Deep Neural Networks in Presence of Noise - Dealing with and Taking Advantage of Distortions | Image classification | data augmentation | ノイズによる分類誤差の検証 シビアな条件のみCNNでBM3Dよりわずかに精度向上 | BM3D | https://www.researchgate.net/profile/Michal_Koziarski/publication/319238759_Image_recognition_with_deep_neural_networks_in_presence_of_noise_-_Dealing_with_and_taking_advantage_of_distortions/links/5a65c2004585158bca52b573/Image-recognition-with-deep-neural-networks-in-presence-of-noise-Dealing-with-and-taking-advantage-of-distortions.pdf | ||||
2019 | Toyota Research Institute | An Attention-based Recurrent Convolutional Network for Vehicle Taillight Recognition | Tail Light Detection | https://arxiv.org/pdf/1906.03683.pdf | |||||||
2016 | Tohoku University | Traffic Light Detection Considering Color Saturation Using In-Vehicle Stereo Camera | TLR | LEDフリッカ 輝度飽和は隣接領域から判定 | https://www.jstage.jst.go.jp/article/ipsjjip/24/2/24_349/_pdf/-char/ja | ||||||
2018 | Shandong University | Robust Lane Detection for Complicated Road Environment Based on Normal Map | Lane Detection | complicated road environment | 勾配推定 ステレオでデプスを出す | http://www.mit.edu/~yanyanxu/doc/ACCESS_Yuan_2018.pdf | |||||
2018 | Mobileye | On a Formal Model of Safe and Scalable Self-driving Cars | ADAS | https://arxiv.org/pdf/1708.06374.pdf | |||||||
2018 | Mobileye | Vision Zero: Can Roadway Accidents be Eliminated without Compromising Traffic Throughput | ADAS | brakeタイミングに関する数式? REMがあれば車線検出が不要とか | https://www.mobileye.com/responsibility-sensitive-safety/vision_zero_with_map.pdf | https://s21.q4cdn.com/600692695/files/doc_presentations/2019/01/Mobileye_CES2019.pdf | |||||
2019 | Tampere University | Architecture Search by Estimation of Network Structure Distributions | NAS | probabilistic representation of a neural network structure | 20GPU days, 77.29%@CIFAR-100 | Shake-Shake [68] | https://arxiv.org/pdf/1908.06886.pdf | ||||
2019 | Xiaomi | SCARLET-NAS: Bridging the gap Between Scalability and Fairness in Neural Architecture Search | NAS | SCARLET (SCAlable supeRnet with Linearly Equivalent Transformation), harmonize the supernet training | 76.9% top-1 accuracy on ImageNet | https://arxiv.org/pdf/1908.06022.pdf | |||||
2019 | Inria, CNRS | Adaptative Inference Cost With Convolutional Neural Mixture Models | Acceleration | Convolutional Neural Mixture Models (CNMMs) | セマセグでの評価もある, CNNのアンサンブル法 | https://arxiv.org/pdf/1908.06694.pdf | |||||
2018 | Cornell University, Fudan University, Tsinghua University, FAIR | MULTI-SCALE DENSE NETWORKS FOR RESOURCE EFFICIENT IMAGE CLASSIFICATION | Acceleration | MSDNet | 下記からの引用, anytime classification, budgeted batch classification | https://arxiv.org/pdf/1703.09844.pdf | |||||
2019 | Tsinghua University, Baidu, University of Oxford | Improved Techniques for Training Adaptive Deep Networks | Acceleration | Adaptive inference | 上記Adaptative Inference Cost ...より精度が高い | MSDNet | https://arxiv.org/pdf/1908.06294.pdf | https://github.com/kalviny/IMTA | |||
2019 | Nanjing University of Science and Technology, Dalian University of Technology, Google, SenseTime, Nankai University | Image Formation Model Guided Deep Image Super-Resolution | Super Resolution | エリアシングのなさで抜きんでてる、シンプルなアルゴ | RDN [35], DBPN [6] | https://arxiv.org/pdf/1908.06444.pdf | |||||
2019 | Chinese Academy of Sciences | Consistent Scale Normalization for Object Recognition | Object Detection | Consistent Scale Normalization (CSN) | 46.5 mAP with a ResNet-101(COCO) | SNIP, SNIPER | https://arxiv.org/pdf/1908.07323.pdf | ||||
2019 | Dalian University of Technology, Adobe | Towards High-Resolution Salient Object Detection | Object Detection | Attended Patch Sampling, Saliency | https://arxiv.org/pdf/1908.07274.pdf | https://github.com/yi94code/HRSOD | |||||
2019 | Texas A&M University, MIT-IBM Watson AI Lab | AutoGAN: Neural Architecture Search for Generative Adversarial Networks | NAS | GANのためのNAS | FID:31.01 | ProbGAN[14] | https://arxiv.org/pdf/1908.03835.pdf | https://github.com/TAMU-VITA/AutoGAN | |||
2018 | A Survey on Shadow Detection Techniques in a Single Image | Shadow removal | 最も充実したサーベイ | http://itc.ktu.lt/index.php/ITC/article/download/15012/9331 | |||||||
2012 | Zhejiang University | Review of Shadow Detection and De-shadowing Methods in Remote Sensing | Shadow removal | 4アプローチのPro/Con | |||||||
2012 | Lovely Professional University | Review on Shadow Detection and Removal Techniques/Algorithms | Shadow removal | http://www.ijcst.com/vol31/3/rajni.pdf | |||||||
2016 | Gachon University | A Review on various widely used shadow detection methods to identify a shadow from images | Shadow removal | 検出のみの簡単なサーベイ | http://www.ijsrp.org/research-paper-0716/ijsrp-p5527.pdf | ||||||
2016 | SHADOW DETECTION AND REMOVAL FROM URBAN HIGH RESOLUTION REMOTE SENSING IMAGES | Shadow removal | HSV, normalized saturation-intensity difference index (NSVDI) | NSVDIは簡単に試せる | http://ena.lp.edu.ua/bitstream/ntb/37386/1/5_42-49.pdf | ||||||
2013 | Shadow Detection and Removal from a Single Image Using LAB Color Space | Shadow removal | Lab | http://www.cit.iit.bas.bg/CIT_2013/v13-1/S_Murali,%20V_Govindan.pdf | |||||||
2013 | Queensland University of Technology | Dealing with Shadows: Capturing Intrinsic Scene Appearance for Image-based Outdoor Localisation | Shadow removal | shadow invariant images | 理論丁寧 | http://www.robots.ox.ac.uk/~mobile/Papers/2013IROS_corke.pdf | |||||
2017 | Eliminating the observer effect: Shadow removal in orthomosaics of the road network | Shadow removal | 車載, BEV | http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w3/Tanathong_Eliminating_the_Observer_ICCV_2017_paper.pdf | |||||||
2009 | REMOVING SHADOWS FROM A SINGLE REAL-WORLD COLOR IMAGE | Shadow removal | log-logの射影がベース | http://www.paper.edu.cn/scholar/showpdf/OUT2YNxIOTT0IxeQh | |||||||
2006 | Shanghai JiaoTong University | Shadow Removal from a Single Image | Shadow removal | separate the vague shadows, color invariant | http://lxu.me/mypapers/XuL_ShadowRemoval.pdf | ||||||
2017 | Chinese Academy of Sciences | DeshadowNet: A Multi-context Embedding Deep Network for Shadow Removal | Shadow removal | DeshadowNet, global localization/appearance/semantic modeling network | 300ms(640 × 480) | http://www.cs.cityu.edu.hk/~rynson/papers/cvpr17a.pdf | |||||
2013 | Wuhan University | Fast Shadow Removal Using Adaptive Multi-scale Illumination Transfer | Shadow removal | interactive brushing, adaptive illumination transfer approach | http://graphvision.whu.edu.cn/papers/EGauthorGuidelines-cgf-fin.pdf | ||||||
2011 | University of Illinois | Single-Image Shadow Detection and Removal using Paired Regions | Shadow removal | Graph-cut Inference | http://dhoiem.cs.illinois.edu/publications/cvpr11_shadow.pdf | ||||||
2007 | University of Haifa | Texture-Preserving Shadow Removal in Color Images Containing Curved | Shadow removal | penumbra scale factors | http://cs.haifa.ac.il/hagit/papers/CONF/CVPR07_ArbelHelOr-shadowRemoval.pdf | ||||||
2019 | The Chinese University of Hong Kong, Chinese Academy of Sciences | Mask-ShadowGAN: Learning to Remove Shadows from Unpaired Data | Shadow removal | CycleGAN | Gong et al. [8] | https://arxiv.org/pdf/1903.10683.pdf | |||||
2015 | The University of Western Australia | Automatic Shadow Detection and Removal from a Single Image | Shadow removal | CNN, CRF, SLIC, SMOTE, GMM, EM, Umbra/Penumbra | https://salman-h-khan.github.io/papers/TPAMI15.pdf | ||||||
2011 | EPFL | REMOVING SHADOWS FROM IMAGES USING COLOR AND NEAR-INFRARED | Shadow removal | NIR, Umbra/Penumbra | https://www.cis.rit.edu/~cnspci/references/dip/nir_removing_shadows/salamati2011.pdf | ||||||
2012 | Google Inc. | Shadow Removal for Aerial Imagery by Information Theoretic Intrinsic Image Analysis | Shadow removal | texture and illumination components, non-parametric kernel-based quadratic entropy formulation, Energy Minimization by PCG | https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37743.pdf | ||||||
2019 | Shandong Technology and Business University | Image Shadow Removal Using End-to-End Deep Convolutional Neural Networks | Shadow removal | RSnet, encoder–decoder network, small refinement network. | |||||||
2006 | University of East Anglia | Simple Shadow Removal | Shadow removal | single scaling factor | https://infoscience.epfl.ch/record/111781/files/piecewise_shadows.pdf | ||||||
2002 | The University of East Anglia | Removing Shadows from Images | Shadow removal | log(R/G) vs log(B/G)で射影する元祖 | https://www.cs.sfu.ca/~mark/ftp/Eccv02/shadowless.pdf | ||||||
2012 | Recovery of Chromaticity Image Free from Shadows via Illumination Invariance | Shadow removal | https://pdfs.semanticscholar.org/ad22/03598d569904697c9c6c9f21728d6e5851a2.pdf | ||||||||
2019 | SenseTime, The Chinese University of Hong Kong | Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks | Image classification | Dynamic Grouping Convolution | cosine learning rate schedule [16] and weight initialization of [6] ResNeXt101 | 43M 79.9 | https://arxiv.org/pdf/1908.05867.pdf | ||||
2019 | Zhejiang University | See Clearer at Night: Towards Robust Nighttime Semantic Segmentation through Day-Night Image Conversion | Semantic Segmentation | CycleGAN, ERF-PSPNet | セマセグの精度が夜間で低下することへの対策 夜画像を生成し、昼と同じラベルを与える BDD Datasetを使用 | https://arxiv.org/pdf/1908.05868.pdf | |||||
2019 | Stanford, NVIDIA | A Delay Metric for Video Object Detection: What Average Precision Fails to Tell | Object Detection | delay metric, Key Frame based Methods | video評価におけるAPの欠点 | NAB Score, CaTDet Delay | https://arxiv.org/pdf/1908.06368.pdf | ||||
2017 | NEC | ACCELERATED RANSAC FOR 2D HOMOGRAPHY ESTIMATION BASED ON GLOBAL BRIGHTNESS CONSISTENCY | RANSAC | GLOBAL BRIGHTNESS CONSISTENCY | シーンによってUSACの方がRANSACに対しx54.6 iter削減 | USAC | https://www.nec.com/en/global/rd/people/docs/icip2017_nakano.pdf | ||||
2008 | The University of North Carolina | A Comparative Analysis of RANSAC Techniques Leading to Adaptive Real-Time Random Sample Consensus | RANSAC | spd-up:~305.7 | PROSAC | http://inf.ethz.ch/personal/pomarc/pubs/RaguramECCV08.pdf | |||||
2019 | the University of Tokyo | How Far Should Self-Driving Cars ‘See’? Effect of Observation Range on Vehicle Self-Localization | SLAM | ERROR AND MATCHING TIMEを評価 Velodyne LiDAR VLP-16, GPS, IMU, and CAN | 8.1cm誤差、39.3 ms matching time | https://arxiv.org/ftp/arxiv/papers/1908/1908.06588.pdf | |||||
2019 | End-to-End Deep Neural Network Design for Short-term Path Planning | Path Planning | https://hal.archives-ouvertes.fr/hal-02266802/document | ||||||||
2019 | Tsinghua University | Real-time lane detection and tracking for autonomous vehicle applications | Lane Detection | model-based clustering | 93.4%, 7ms(i7-6820EQ, 768x480) | https://journals.sagepub.com/doi/pdf/10.1177/0954407019866989 | |||||
2001 | Universit´e Blaise Pascal | A model-driven approach for real-time road recognition | Lane Detection | confidenceの算出方法参考 | https://www.researchgate.net/publication/220465009_A_model-driven_approach_for_real-time_road_recognition | ||||||
2004 | Universitat Autonoma de Barcelona, Volkswagen | Detection of Lane Markings based on Ridgeness and RANSAC | Lane Detection | ridgeness measure, RANSAC | 円弧カーブの投影モデルをフィッティング カーブに強い | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.135.7283&rep=rep1&type=pdf | |||||
2009 | Georgia Institute of Technology | ROBUST LANE DETECTION AND TRACKING WITH RANSAC AND KALMAN FILTER | Lane Detection | RANSAC | Kalman filterでラインの位置と方向、それらの速度を追跡 | ||||||
2018 | Yeungnam University | Lane Detection Using Labeling Based RANSAC Algorithm | Lane Detection | Canny, IPM, Morphology, K-means, RANSAC | Gauss-Newtonで2次多項式係数を求めるlane fitting | https://waset.org/publications/10008894/lane-detection-using-labeling-based-ransac-algorithm | |||||
2014 | California Institute of Technology | Real time Detection of Lane Markers in Urban Streets | Lane Detection | Hough Transform, RANSAC Spline fitting | IPMの行列が記載されている、遠くの破線・カーブが苦手そう Homography参考 | https://arxiv.org/pdf/1411.7113.pdf | |||||
2019 | NY | InverseRenderNet: Learning single image inverse rendering | Inverse Rendering | CNNでalbedo/normalを出す | https://arxiv.org/pdf/1811.12328.pdf | ||||||
2016 | University of Technology Sydney | Deep Neural Network for Structural Prediction and Lane Detection in Traffic Scene | Lane Detection | IPM, RANSAC | SVM<CNN |
http://islab.ulsan.ac.kr/files/announcement/552/Deep%20Neural%20Network%20for%20Structural%20Prediction%20and%20lane%20detection%20in%20traffic%20scene.pdf | |||||
2015 | Campus Universitario de Santiago | Multimodal inverse perspective mapping | IPM | multimodal sensor fusion, laser range finder, Inverse projection | LiDARで障害物を除去したIPM(using a laser sensor to detect mappable regions) | http://www.cvc.uab.es/~asappa/publications/J__Elsevier_IF_Vol_24_July_2015_pp_108-121.pdf | |||||
2019 | Zhejiang University | A Multimodal Vision Sensor for Autonomous Driving | Stereo, Polarization, Panoramic | Sony IMX250MZR, 45degずつ4偏光、DOPで水たまり検知 | https://arxiv.org/pdf/1908.05649.pdf | ||||||
2019 | Huazhong University of Science and Technology | IoU-balanced Loss Functions for Single-stage Object Detection | Object Detection | IoU-balanced classification loss | improve AP by 1.1%(RetinaNet ResNet50) | https://arxiv.org/ftp/arxiv/papers/1908/1908.05641.pdf | |||||
2019 | Shanghai Jiao Tong University, Huawei | R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object | Object Detection | Remote sensing, Rotation detection | https://arxiv.org/pdf/1908.05612.pdf | ||||||
Einconv: Exploring Unexplored Tensor Decompositions for Convolutional Neural Networks | |||||||||||
2016 | Nanyang Technological University | DEPTH OF FIELD GUIDED REFLECTION REMOVAL | Highlight Removal | Multi-scale Background Edge Extraction | 影分離に応用できないか | https://rose.ntu.edu.sg/Publications/Documents/Reflection%20Removal/Depth%20of%20field%20guided%20reflection%20removal.pdf | |||||
2010 | University of Illinos | Real-time Specular Highlight Removal Using Bilateral Filtering | Highlight Removal | http://vision.ai.illinois.edu/publications/eccv-10-qingxiong-yang.pdf | |||||||
2015 | Xi’an Jiaotong University | Saturation-preserving Specular Reflection Separation | Highlight Removal | dichromatic reflection model | Matlab code on a 3.4GHz Intel Core i7 cpu is 46.8s | https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Liu_Saturation-Preserving_Specular_Reflection_2015_CVPR_paper.pdf | |||||
2017 | EPFL | Single Image Reflection Suppression | Reflection Suppression | Laplacian data fidelity term and an l0 gradient sparsity term | 反射像の低減にとどまる | Li and Brown [12] Wan et al. [22] | https://infoscience.epfl.ch/record/227363/files/2017-CVPR-Arvanitopoulos.pdf; | ||||
2003 | The University of Tokyo | Separating Reflection Components of Textured Surfaces using a Single Image | Highlight Removal | chromaticity intensity space, intensity logarithmic differentiation | http://ljk.imag.fr/membres/Bill.Triggs/events/iccv03/cdrom/iccv03/0870_tan.pdf | ||||||
2011 | Cyprus Institute CaSToRC | A Survey of Specularity Removal Methods | Highlight Removal | Dichromatic Reflectance Model | DoPでspecularを分離(拡散反射は無偏光) | http://vcg.isti.cnr.it/Publications/2011/ABC11/j.1467-8659.2011.01971.x.pdf | |||||
2015 | ? | Fast and High Quality Highlight Removal from A Single Image | Highlight Removal | L2 NORMALIZED DICHROMATIC MODEL | 2色性反射モデルを一般化 | STAR | https://arxiv.org/pdf/1512.00237.pdf | ||||
2015 | Sogang University | BREN: Body Reflection Essence-Neuter Model for Separation of Reflection Components | Highlight Removal | dichromatic reection | Wikipediaの2色性反射モデルのページで引用 | maximum chromaticity-based methods | https://www.researchgate.net/publication/275056779_BREN_Body_Reflection_Essence-Neuter_Model_for_Separation_of_Reflection_Components/link/5554812b08aeaaff3bf1dcab/download | ||||
2015 | Chinese Academy of Sciences | An Image Highlights Removal Method with Polarization Principle | Highlight Removal | 下と著者は異なる | https://pdfs.semanticscholar.org/318d/5ccabbf3047130ef2222dc76530e76220c88.pdf | ||||||
2016 | Chinese Academy of Sciences | A Method of Removing Reflected Highlight on Images Based on Polarimetric Imaging | Highlight Removal | SSS HPのデモで引用されていた Brewster角近傍でなくても偏光で反射を低減可能 (無偏光の屈折成分の抽出) DOLPは4つの偏光角の反射強度を使う DOLPは反射角度に依存するために、法線が可視化できる | http://downloads.hindawi.com/journals/js/2016/9537320.pdf | https://www.sony-semicon.co.jp/products_ja/IS/sensor5/index.html | |||||
2008 | Tandent Vision Science, Inc. | US 8,319,821 B2 POLARIZATION-BASED SHADOW DETECTION | Polarimetric imaging | https://patentimages.storage.googleapis.com/70/4e/4d/fe569a5804d9e2/US8319821.pdf | |||||||
2006 | University of Pennsylvania | Separation and contrast enhancement of overlapping cast shadow components using polarization | Polarimetric imaging | degree-of-polarization image | 偏光による影分離の先行例 特許の参考、影と偏光の関係について述べているので理論の根拠として使えるか DoP定義式 0, 45, 90degでlinear polarizerを回転 DoP画像を見ても影除去・検出はできていない 唯一価値がありそうな知見は、“shadow” cast by the much weaker fluorescent light from the right hand side is revealed | https://www.osapublishing.org/oe/abstract.cfm?uri=OE-14-16-7099 | |||||
2019 | Stanford University | Segmenting the Future | Semantic Segmentation | LSTMではなく3D conv 複数フレーム入力→Encoder→Conv3D Forecasting Module→Decoder motionなしより高精度 OSSは準備中 | mIoU:40.81(Long-term)~65.08(Short-term) | XS2XS, ConvLSTM | https://arxiv.org/pdf/1904.10666.pdf | https://github.com/eddyhkchiu/segmenting_the_future | |||
Intelligent Automation Inc | Stereo Imaging with Uncalibrated Camera | Stereo | F行列推定し画像変換 | https://www.researchgate.net/profile/Chiman_Kwan/publication/220844935_Stereo_Imaging_with_Uncalibrated_Camera/links/563a326d08ae337ef2983d69/Stereo-Imaging-with-Uncalibrated-Camera.pdf | |||||||
2012 | CMU | Online Continuous Stereo Extrinsic Parameter Estimation | Camera Calibration | 振動等の外乱からextrinsicはonlineが必要 | https://kilthub.cmu.edu/articles/Online_Continuous_Stereo_Extrinsic_Parameter_Estimation/6557093/files/12039368.pdf | ||||||
2012 | Xi’an University of Technology | A New Method of Stereo Localization Using Dual-PTZ-Cameras | Rectification | MSER, SIFT | intrinsic/extrinsic parameter両方必要 異焦点の実験では、ズーム補正(焦点距離比による縮小)が前処理として必要 SIFTにMSERを組み合わせることでマッチングが大幅に低減 | https://www.researchgate.net/profile/Xin_Jing8/publication/262168457_A_New_Method_of_Stereo_Localization_Using_Dual-PTZ-Cameras/links/554316b20cf23ff716837b86/A-New-Method-of-Stereo-Localization-Using-Dual-PTZ-Cameras.pdf | |||||
2018 | Harbin Engineering University | Automatic Rectification of the Hybrid Stereo Vision System | Rectification | ASIFT | 360degカメラと通常カメラのステレオ | https://www.researchgate.net/publication/328150741_Automatic_Rectification_of_the_Hybrid_Stereo_Vision_System | |||||
1999 | Microsoft Research | Computing Rectifying Homographies for Stereo Vision | Rectification | Similarity Transform | 平行化に伴う歪の最小化, 平行化に加えshear? | http://dev.ipol.im/~morel/Dossier_MVA_2011_Cours_Transparents_Documents/2011_Cours7_Document2_Loop-Zhang-CVPR1999.pdf | |||||
2009 | Czech Technical University | Epipolar Rectification for Stereovision | Rectification | omnidirectional camera, spectral loss | State of the artのまとめが参考になる | Geyer and Daniilidis, Pollefeys’ | |||||
2011 | Gwangju Institute of Science and Technology | An efficient image rectification method for parallel multi-camera arrangement | Rectification | 有料のため読んでない | https://ieeexplore.ieee.org/abstract/document/6018853 | ||||||
1999 | Heriot-Watt University | Projective Rectification without Epipolar Geometry | Rectification | homographies, Levenberg-Marquardt algorithm, hierarchical rectification | 普通の再投影誤差(L2)最小化 疑似コードあり | https://www.researchgate.net/profile/Francesco_Isgro/publication/3813442_Projective_rectification_without_epipolar_geometry/links/09e4150b20494a205e000000.pdf | |||||
2013 | Mitsubishi Electric Research Laboratories, Inc. | Disparity Estimation of Misaligned Images in a Scanline Optimization Framework | Stereo | Cross-check, semi-global matching | http://www.merl.com/publications/docs/TR2013-025.pdf | ||||||
2012 | CSIRO Mathematics | Closed-form Stereo Image Rectification | Rectification | uncalibrated stereo, keystone e↵ect correction, fundamental matrix | distortion軽減 | Mallon and Whelan’s method | http://vision-cdc.csiro.au/changs/doc/sun12ivcnz.pdf | http://vision-cdc.csiro.au/rectify2v/ | |||
1999 | K.U.Leuven | A simple and efficient rectification method for general motion | Rectification | cylindrical rectification, fundamental matrix | epipole中心の極座標表現 | https://www.cs.unc.edu/~marc/pubs/PollefeysICCV99.pdf | |||||
2019 | CASIA, University of Oxford | Fast Online Object Tracking and Segmentation: A Unifying Approach | Tracking | Siamese approach, video object segmentation | ECO | https://arxiv.org/pdf/1812.05050.pdf | http://www.robots.ox.ac.uk/~qwang/SiamMask/ | ||||
2019 | NAVER Corp. | EXTD: Extremely Tiny Face Detector via Iterative Filter Reuse | Object Detection | 数層の浅い軽量なbackbone network(MobileNet v2のInverted Residual Block)を繰り返し再帰的にかける SSD-likeなものとFPN-likeなものを提案 PReLU (またはLeakyReLU) の方が良い | #Params1/200でmAP-0.3 | https://arxiv.org/pdf/1906.06579.pdf | https://uiiurz1.hatenablog.com/entry/2019/06/24/221506 | ||||
2019 | University of Chinese Academy of Sciences, University of Oxford, Huawei | CenterNet: Keypoint Triplets for Object Detection | Object Detection | CenterNet | CornerとCenterの3点のembedding情報のみをpoolingするため、ROI poolingのように計算量が大きくならない CornerとCenterを別ブランチで推定 | mAP47.0, 340ms/43.5, 270ms | CornerNet | https://engineer.dena.jp/2019/07/cv-papers-19-keypoint-object-detection.html | |||
2019 | UT Austin, UC Berkeley | Objects as Points | Object Detection | CenterNet, Deep Layer Aggregation (DLA) | Real Time Object DetectionのSOTA 中心のみをヒートマップで予測 アンカーが存在せず、bounding boxサイズを直接、クラス毎に出力 deformable convolutionレイヤをupsampling部に用いている 向きを回帰して3D BBox推定を行う CornerNetの律速はcombinatorial grouping stage(keypointのペア探索) classごとにFmapのピークを100個検出 | mAP45.1, 7.8fps/41.6, 28fps(CornerNetと同精度で7倍速) | CornerNet | https://arxiv.org/pdf/1904.07850v2.pdf | https://github.com/xingyizhou/CenterNet | https://paperswithcode.com/task/real-time-object-detection https://www.slideshare.net/DeepLearningJP2016/dlobjects-as-points | |
2009 | University of Udine | Stereo Rectification of Un-calibrated and Heterogeneous Image-Pairs | Rectification | different focal lengths and/or image resolutions, Levenberg-Marquardt algorithm | Fusielloらと異なり、カメラパラを求めない直接平行化 焦点距離比に基づきゼロ埋めと縮小で同一サイズにする 平行化として、SIFTマッチングペアについてm’H’FHmが最小となるH’,Hを求める 異焦点対応だが平行化が面倒(LM法) | https://pdfs.semanticscholar.org/2afe/d3c5397736b422050496c501f70d63d2ac50.pdf | |||||
2017 | Kyungpook National University | Calibration of a Different Field-of-view Stereo Camera System using an Embedded Checkerboard Pattern | Rectification | 特殊なチェスボード FOV大小のステレオという特許と同じ設定 式が全然出てこない(Fusiello2000を使っているらしい) | https://www.scitepress.org/papers/2017/62678/62678.pdf | ||||||
2013 | Paris-Est Marne-la-Vallee University | Camera array image rectification and calibration for stereoscopic and autostereoscopic displays | Rectification | homography, LevenbergMarquartd method, quasi Euclidean | 多眼への拡張 解像度ばらつきにも対応 | https://hal-upec-upem.archives-ouvertes.fr/hal-00908586/document | |||||
2009 | National Tsing Hua University | Self Image Rectification for Uncalibrated Stereo Video with Varying Camera Motions and Zooming Effects | Rectification | 2.5fps | http://www.mva-org.jp/Proceedings/2009CD/papers/02-03.pdf | ||||||
2000 | Universita di Verona | A Compact Algorithm for Rectification of Stereo Pairs | Rectification | 本文中にMATLAB code 異焦点で使えることが2017Calibration of a Different ...で言及されている 光学中心cを求め、 誤差はノイズσ0.025までは線形増大 エピポーラ線算出不要だがtwo vector cross multiplicationsにより計算量増大 | http://www.diegm.uniud.it/fusiello/papers/mva99.pdf | https://webcourse.cs.technion.ac.il/236873/Winter2015-2016/ho/WCFiles/Rectification.pdf | |||||
2011 | Fuzhou University | Stereo rectification of calibrated image pairs based on geometric transformation | Rectification | 先行研究のまとめが参考になる two virtual camerasが新しい y,z軸周りの回転で平行化 座標変換と再投影の2ステップ 途中から焦点距離が同一という仮定が入る | https://pdfs.semanticscholar.org/6bcf/f783d1171f3ee3f6a4a8f8ef96533cab48e2.pdf | ||||||
2016 | UC Berkeley, Nokia | Depth from Semi-Calibrated Stereo and Defocus | Stereo | 補助ステレオカメラでデプス Fusielloより平行化高精度 | https://zpascal.net/cvpr2016/Wang_Depth_From_Semi-Calibrated_CVPR_2016_paper.pdf | ||||||
2010 | Paris-Est/École des Ponts ParisTech | Three-step image rectification | Rectification | Fusielloが一番いい? | https://hal-enpc.archives-ouvertes.fr/file/index/docid/654415/filename/BMVC10.pdf | ||||||
2011 | Universit´e Paris-Est | Quasi-Euclidean Epipolar Rectification | Rectification | Fusielloの制約(主点の固定)をx方向だけ自由に MissStereo動作確認済み | http://www.ipol.im/pub/art/2011/m_qer/article_lr.pdf | http://www.ipol.im/pub/art/2011/m_qer/?utm_source=doi | |||||
2008 | Quasi-Euclidean Uncalibrated Epipolar Rectification | Rectification | MATLAB code | http://www.diegm.uniud.it/fusiello/papers/qeur.pdf | http://www.diegm.uniud.it/fusiello/demo/rect/ | ||||||
2011 | Quasi-Euclidean epipolar rectification of uncalibrated images | Rectification | uncalibrated stereo | MATLAB code 未較正の場合 Sampson誤差近似の最小化 | https://core.ac.uk/download/pdf/53327714.pdf | http://www.diegm.uniud.it/fusiello/demo/rect/ | https://webcourse.cs.technion.ac.il/236873/Winter2015-2016/ho/WCFiles/Rectification.pdf | ||||
2016 | Pedestrian lane detection in unstructured scenes for assistive navigation | Lane Detection | Vanishing point estimation, color tensor | 路面標示なし 消失点推定参考 最高速度・精度 2階微分から求めたdominant local orientationに基づき、各エッジ画素からVP候補にvoting(これが重いのか) | 595ms | https://reader.elsevier.com/reader/sd/pii/S1077314216000369?token=CBF4142865CCB3AB8088BE1684620961065C9A71DDD4DBC6596D9DCA535FBE5DAF65D43EACBED30D98AF2162D708E171 | https://documents.uow.edu.au/~phung/plvp_dataset.html | ||||
2011 | EDLines: A real-time line segment detector with a false detection control | 下記からの引用 LSDの11倍速い | http://c-viz.eskisehir.edu.tr/pdfs/EDLines2011ICIP.pdf | https://github.com/shaojunluo/EDLinePython https://github.com/CihanTopal/ED_Lib | |||||||
2018 | Karabuk University | A Study on Real-Time Detection Method of Lane and Vehicle for LaneChange Assistant System Using Vision System on Highway | Lane Detection | EDLines, Vanishing point estimation | 43ms | https://reader.elsevier.com/reader/sd/pii/S2215098617317317?token=E6F75C83EFC772C7D47095F6429B912600486F11E89AB23B8B5084EBC15A18C6A373E027203468BC1C42010C25ED336D | |||||
2008 | Technion | Active Polarization Descattering | Dehaze | https://www.researchgate.net/profile/Tali_Treibitz/publication/23791813_Active_Polarization_Descattering/links/02e7e532ac0af505b2000000.pdf | |||||||
2003 | Columbia University | Polarization-based vision through haze | Dehaze | http://webee.technion.ac.il/~yoav/publications/haze_osa.pdf | |||||||
2018 | University of Oxford | Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment | SLAM | 物体サイズのprior利用したscale推定 | https://ora.ox.ac.uk/objects/uuid:74aaa7b5-5b14-4feb-b672-fa1802a804c6/download_file?file_format=pdf&safe_filename=Frost%2Bet%2Bal.pdf&type_of_work=Journal+article | ||||||
2004 | Indian Institute of Technology | Depth Estimation and Image Restoration Using Defocused Stereo Pairs | DFD, stereo | http://www.ee.iitm.ac.in/~raju/journals/j14.pdf | |||||||
2010 | University of Udine | Stereo rectification of uncalibrated and heterogeneous images | image shrinking, rectification, SIFT | still&video(w/ zoom)のペア x’TH’TF∞Hx最小化による平行化変換(homography)推定 | https://pdfs.semanticscholar.org/2afe/d3c5397736b422050496c501f70d63d2ac50.pdf | ||||||
2000 | INRIA | Matching Images with Different Resolutions | scale-space representatio | https://hal.inria.fr/inria-00548297/document | |||||||
1999 | The University of Readin | Critical Motion Sequences for the Self-Calibration of Cameras and Stereo Systems with Variable Focal Length | Camera Calibration | Potential Absolute Conics | https://hal.inria.fr/inria-00525676/document | ||||||
Method for measuring stereo camera depth accuracy based on stereoscopic vision | Stereo | depth resolution参考 距離の2乗に比例 | https://www.researchgate.net/profile/Mikko_Kytoe/publication/235349563_Method_for_measuring_stereo_camera_depth_accuracy_based_on_stereoscopic_vision/links/00b4951af44a7e17f5000000/Method-for-measuring-stereo-camera-depth-accuracy-based-on-stereoscopic-vision.pdf | ||||||||
2006 | University Siegen | Fusion of Stereo-Camera and PMD-Camera Data for Real-Time Suited Precise 3D Environment Reconstruction | Stereo | “Winner Takes It All” and ”Simulated Annealing” stereo matching algorithm | 異焦点ステレオの視差-距離変換式(2) 平行化の言及なし | https://www.researchgate.net/profile/Martin_Stommel/publication/221063819_Fusion_of_Stereo-Camera_and_PMD-Camera_Data_for_Real-Time_Suited_Precise_3D_Environment_Reconstruction/links/0a85e53435f1826115000000/Fusion-of-Stereo-Camera-and-PMD-Camera-Data-for-Real-Time-Suited-Precise-3D-Environment-Reconstruction.pdf#page=7 | |||||
Katholieke Universiteit Leuven | Euclidean 3D reconstruction from image sequences with variable focal lengths | Camera Calibration | 異なる焦点距離の推定がテーマ | https://www.researchgate.net/profile/Marc_Proesmans/publication/2882771_Euclidean_3D_reconstruction_from_image_sequences_with_variable_focal_lengths/links/53ed13820cf26b9b7dc14646.pdf | |||||||
1990 | Using Vanishing Points for Camera Calibration | Camera Calibration | http://www.close-range.com/docs/Using_vanishing_points_for_camera_calibration--Caprile-Torre.pdf | ||||||||
1999 | The University of Readin | Critical Motion Sequences for the Self-Calibration of Cameras and Stereo Systems with Variable Focal Length | Stereo | Variable Focal Length | 特許の参考 | https://hal.inria.fr/inria-00525676/document | |||||
2019 | Tencent | Few-Shot Object Detection with Attention-RPN and Multi-Relation Detecto | Object Detection | Attention-Based Region Proposal Network, Few-Shot | 恐ろしく少ない学習で認識できる online学習ができる? 60000iterations | LSTD | https://arxiv.org/pdf/1908.01998.pdf | https://github.com/fanq15/Few-Shot-Object-Detection-Dataset | |||
2012 | CMLA, ENS Cachan | LSD: a Line Segment Detector | LSD | https://www.ipol.im/pub/art/2012/gjmr-lsd/article.pdf | http://www.ipol.im/pub/art/2012/gjmr-lsd/?utm_source=doi | ||||||
2019 | POLITECNICO DI TORINO | Design and Simulation of Autonomous Driving Algorithms | Lane Detection | Canny, Hough, Kalman Filter | Pythonコード付きのM論文, IPMの式(4.2.3), 車の運動モデルの式が詳しい | https://webthesis.biblio.polito.it/11650/1/tesi.pdf | |||||
2016 | Moscow Institue of Physics and Technology, MIT Media Lab | Shape from Mixed Polarization | SfP | Specular+Diffuse Reflection | 特許の参考 | https://arxiv.org/pdf/1605.02066.pdf | |||||
2019 | University of Chinese Academy of Sciences, TuSimple | Revisiting Feature Alignment for One-stage Object Detection | Object Detection | RoIConv | im2colとRoIAlignの関係を分析 RetinaNetが + RoIConvでAP5%向上 | CenterNet | https://arxiv.org/pdf/1908.01570.pdf | ||||
2018 | UC Berkeley, Facebook | MIXED PRECISION QUANTIZATION OF CONVNETS VIA DIFFERENTIABLE NEURAL ARCHITECTURE SEARCH | NAS | DNAS, GumbelSoftmax | lossに#PARAM/#FLOPsを入れる | https://arxiv.org/pdf/1812.00090.pdf | |||||
2018 | UC Berkeley, Princeton University, Facebook | FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search | NAS | Operator Latency LUT | lossにlatencyを入れる | https://arxiv.org/pdf/1812.03443.pdf | https://github.com/facebookresearch/mobile-vision | https://qiita.com/cvusk/items/e7c9bb30c801996cd973 | |||
2019 | DeepScale | SqueezeNAS: Fast neural architecture search for faster semantic segmentation | Semantic Segmentation, NAS | MAC/Latency-aware Search | Fbnetに近い | 34.57ms(Cityscapes), MobileNetV3と同等, 7~14.6GPU days | https://arxiv.org/pdf/1908.01748.pdf | ||||
2019 | Bosch | Group Pruning using a Bounded-`p norm for Group Gating and Regularization | Model Compression | Channel level sparsity | https://lmb.informatik.uni-freiburg.de/Publications/2019/Bro19a/mummadi_gcpr19.pdf | ||||||
2019 | Carnegie Mellon University | Random Search and Reproducibility for Neural Architecture Search | NAS | Random Search with Weight Sharing | 既出 | https://arxiv.org/pdf/1902.07638.pdf | https://github.com/liamcli/randomNAS_release | ||||
2019 | Hong Kong Baptist University | AutoML: A Survey of the State-of-the-Art | AutoML | サーベイ | |||||||
2016 | University of North Carolina | Structure-from-Motion Revisited | Visual localization | https://demuc.de/papers/schoenberger2016sfm.pdf | |||||||
2019 | TU Munich | To Learn or Not to Learn: Visual Localization from Essential Matrices | Visual localization | DSAC++, RelaPoseNet, SIFT + 5-Point | https://arxiv.org/pdf/1908.01293.pdf | ||||||
2015 | The University of Western Australia | Automatic Shadow Detection and Removal from a Single Image | Shadow removal | Bayesian, matting | 曲面でうまくいかないことが課題 | https://salman-h-khan.github.io/papers/TPAMI15.pdf | |||||
2017 | Peking University, Dalian University of Technology | ORGB: OFFSET CORRECTION IN RGB COLOR SPACE FOR ILLUMINATION-ROBUST IMAGE PROCESSING | Shadow removal | 影対策要注目、1-histeq(U,V)が有効 R vs Gの散布図が原点を通るようなオフセットを加える | http://150.162.46.34:8080/icassp2017/pdfs/0001557.pdf | ||||||
2013 | Yonsei University | Gradient-Enhancing Conversion for Illumination-Robust Lane Detection | Lane Detection | adaptive Canny(LDAで閾値決める) | 色変換を次元圧縮とみなし Hough+モデルフィッティング(遠方だけ?2次) | 96% | IPM+grad-based | http://diml.yonsei.ac.kr/papers/Gradient-Enhancing%20Conversion%20for%20Illumination-Robust%20Lane%20Detection.pdf | |||
2018 | A Robust Lane Detection Method Based on Vanishing Point Estimation | Lane Detection | LSD | 白線のコントラストのためにR+G-Bを使用 scanでnoise filtering VP制約;線分中心の座標と角度からVPを通る直線との距離を算出 | Correct detection rate:92.34%~ | https://reader.elsevier.com/reader/sd/pii/S1877050918305489?token=8DAADEF6CAFCB06D8EF088EEAB7DD10425B2D2E5CDADA0AB59BDF5305590AB8DDC440D5658D8843D6F81F9B7F6AF724F | |||||
2019 | Shandong University, Singapore University of Technology and Design | Road Context-aware Intrusion Detection System for Autonomous Cars | Autonomous Driving | Road context-aware IDS (RAIDS), intrusion detection systems (IDSs), CAN | 520ms以下が必須, 24+64層のCNNで画像の特徴量を生成しCANと合わせて2nd NNに入力し異常判定 | https://arxiv.org/pdf/1908.00732.pdf | |||||
2018 | Beijing University of Technology | Efficient Traffic-Sign Recognition with Scale-aware CNN | TSR | OHEM, "Inception" module | 356ms(K40, 960×1280), AP:85.19%(small, STSD) | Faster R-CNN, RPN | https://arxiv.org/pdf/1805.12289.pdf | ||||
2019 | Tufts University, CMU, IBM | Ablate, Variate, and Contemplate: Visual Analytics for Discovering Neural Architectures | NAS | Rapid Exploration of Model Architectures and Parameters | コンパクトなベースラインを人手で探索してそれを最適化 | https://arxiv.org/pdf/1908.00387.pdf | https://github.com/dylancashman/remap_nas http://www.eecs.tufts.edu/~dcashm01/snacs/ | ||||
2017 | Speed/accuracy trade-offs for modern convolutional object detectors | Object Detection | http://openaccess.thecvf.com/content_cvpr_2017/papers/Huang_SpeedAccuracy_Trade-Offs_for_CVPR_2017_paper.pdf | ||||||||
2017 | Cornell University | Reconfiguring the Imaging Pipeline for Computer Vision | RAW認識。Bayer化simができるCRIP。解像度1/4でも分類精度低下は1%未満。Bayerの疑似3ch化。 | https://github.com/cucapra/approx-vision | |||||||
2016 | The University of Western Australia | Automatic Shadow Detection and Removal from a Single Image | Shadow removal | DL応用の | https://salman-h-khan.github.io/papers/TPAMI15.pdf | ||||||
2018 | University of Amsterdam | Three for one and one for three: Flow, Segmentation, and Surface Normals | Segmentation, Flow, and Surface Normals | https://arxiv.org/pdf/1807.07473.pdf | |||||||
2019 | University of Southern California | Every Pixel Counts ++: Joint Learning of Geometry and Motion with 3D Holistic Understanding | Monocular Depth Estimation | Odometry | Motion/OF/Depthを別のNNで同時に出す | Abs Rel:0.141(Eigen test split) | https://arxiv.org/pdf/1810.06125.pdf | https://github.com/chenxuluo/EPC | |||
2016 | Australian National University | Reliable Scale Estimation and Correction for Monocular Visual Odometry | Monocular Visual Odometry | absolute scale | スケール推定参考, Homography→pose→ camera height推定→scale補正 | https://www.researchgate.net/profile/Dingfu_Zhou/publication/305257884_Reliable_Scale_Estimation_and_Correction_for_Monocular_Visual_Odometry/links/5bb9db8ca6fdcc9552d569c2/Reliable-Scale-Estimation-and-Correction-for-Monocular-Visual-Odometry.pdf | |||||
2019 | Peking University | Multi-level Domain Adaptive learning for Cross-Domain Detection | Object Detection, Domain adaptation | Image/Instance-Stage Adaptation, Multi-level Patch-based Loss | https://arxiv.org/pdf/1907.11484.pdf | ||||||
2019 | University of Salerno, University of Granada | DEEP LEARNING IN VIDEO MULTI-OBJECT TRACKING: A SURVEY | Tracking | DLベースのサーベイ | https://arxiv.org/pdf/1907.12740.pdf | ||||||
2017 | SOFT-SLAM: Computationally Efficient Stereo Visual SLAM for Autonomous UAVs | SLAM | KITTIでLiDAR使わないSLAMの最上位, stereo | https://pdfs.semanticscholar.org/df31/78e77fa6655bbcae65c00b732ef240d99fa5.pdf | |||||||
2019 | Uber | Multi-Task Multi-Sensor Fusion for 3D Object Detection | Object Detection | AVOD | http://openaccess.thecvf.com/content_CVPR_2019/papers/Liang_Multi-Task_Multi-Sensor_Fusion_for_3D_Object_Detection_CVPR_2019_paper.pdf | ||||||
2019 | In Defense of Pre-trained ImageNet Architectures for Real-time Semantic Segmentation of Road-driving Images | Semantic Segmentation | SwiftNet | Real-time SSのSOTA | 75.5% MIoU and achieves 39.9 Hz on 1024×2048 images on GTX1080Ti. | https://arxiv.org/pdf/1903.08469v2.pdf | https://paperswithcode.com/sota/real-time-semantic-segmentation-on-cityscapes | ||||
2017 | University of Amsterdam | Modeling Relational Data with Graph Convolutional Networks | Graph Convolutional Network | 下記で使われているGCNの最初の論文 | https://arxiv.org/pdf/1703.06103.pdf | https://qiita.com/tktktks10/items/98d21133cf3e121676c3 | |||||
2019 | Zhejiang University | A Comparative Study of High-Recall Real-Time Semantic Segmentation Based on Swift Factorized Network | Semantic Segmentation | IAL, SwiftNet | U-Netライク、decorderに工夫、t-SNE解析あり | MIOU:72.2, 58.6FPS(1080Ti, CamVid, 2048 × 1024) | Global convolution Networks (GCNet) | https://arxiv.org/pdf/1907.11394.pdf | https://github.com/Katexiang/swiftnet/tree/master/Swift_Factorized_Network(SFN) | ||
2019 | Carnegie Mellon University | Memory- and Communication-Aware Model Compression for Distributed Deep Learning Inference on IoT | Model Compression, Distillation | Network of Neural Networks (NoNN) | IoT想定で複数studentに分散処理(最終層で統合) | x12高速化 | https://arxiv.org/pdf/1907.11804.pdf | ||||
2019 | Fraunhofer Heinrich Hertz Institute | DeepCABAC: A Universal Compression Algorithm for Deep Neural Networks | Model Compression | Context-based Adaptive Binary Arithmetic Coder (CABAC) | 符号化によるモデルサイズ圧縮。速度への言及なし。 | 精度不変で0.3%のサイズ(VGG16) | https://arxiv.org/pdf/1907.11900.pdf | ||||
2019 | Huawei | Learning Instance-wise Sparsity for Accelerating Deep Models | Model Compression, Image classification | instance-wise sparsity | 正則化の変更(空間L2+ch/Layer L1) | prune ratio:46.9(VGG16) | https://arxiv.org/pdf/1907.11840.pdf | ||||
2018 | ETH | Domain Adaptive Faster R-CNN for Object Detection in the Wild | Object Detection | cross-domain robustness, gradient reverse layer (GRL) [15] | cls,/reg.とROIpool前とを後段に回して整合性正則化 | ||||||
2019 | Foveated image processing for faster object detection and recognition In embedded systems using deep convolutional neural networks | from 3.59 FPS to 15.24 FPS, 92.0% of the baseline performance | 取り寄せ中 | ||||||||
2019 | Technical University of Munich | MultiDepth: Single-Image Depth Estimation via Multi-Task Regression and Classification | Monocular Depth Estimation | Multi-Task | cls+reg loss | SILog:16.05 | https://arxiv.org/pdf/1907.11111.pdf | ||||
2019 | Hanyang University | From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation | Monocular Depth Estimation | BTS | KITTIでtop | SILog:11.67 | https://arxiv.org/pdf/1907.10326.pdf | ||||
2018 | Indian Institute of Technology | Digital Foveation: An Energy-Aware Machine Vision Framework | 消費電力76.5%減 | ||||||||
2019 | Hanyang University | Evidence Filter of Semantic Segmented Image From Around View Monitor in Automated Parking System | Semantic Segmentation | GraphSLAM, Around View Monitor, evidence filter | 自動駐車参考, セマセグの課題を解決, 白線・FS・障害物の3クラス | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8758815 | |||||
2017 | Weizmann Institute of Science | On Detection of Faint Edges in Noisy Images | Edge detection | https://arxiv.org/pdf/1706.07717.pdf | |||||||
2018 | Motilal Nehru National Institute of Technology Allahabad | ROAD SURFACE DETECTION FROM MOBILE LIDAR DATA | Road Surface | LiDAR, PCA | Precision:98.85%, 24 minutes(MATLAB) | https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/IV-5/95/2018/isprs-annals-IV-5-95-2018.pdf | |||||
2019 | Chongqing University | An Efficient Compressive Convolutional Network for Unified Object Detection and Image Compression | Compressive Sensing | https://pdfs.semanticscholar.org/76ea/ce51586ad51044aeacb50ce254099eca647f.pdf | |||||||
2019 | Universidade Federal do Esp´ırito Santo | Cross-Domain Car Detection Using Unsupervised Image-to-Image Translation: From Day to Night | Object Detection | CycleGAN, Fake-night images | ELASの研究室?昼の画像だけから夜のラベルなし画像を用いた学習を行う | fake-night+day:86.6% Night:92.0% | https://arxiv.org/pdf/1907.08719.pdf | ||||
2019 | Seoul National Universit | ROI-Based LiDAR Sampling Algorithm in on-Road Environment for Autonomous Driving | Object Detection | LiDAR | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8755965 | ||||||
2019 | University of Chinese Academy of Sciences | MCF3D: Multi-Stage Complementary Fusion for Multi-Sensor 3D Object Detection | Road Marking RecognitionD Object Detection | LIDAR | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8756006 | ||||||
2018 | Beihang University | Receptive Field Block Net for Accurate and Fast Object Detection | Object Detection | 下記からの引用、速度精度両立 | <50ms, SSDと同速度で5%向上(mAP on MS COCO) | http://openaccess.thecvf.com/content_ECCV_2018/papers/Songtao_Liu_Receptive_Field_Block_ECCV_2018_paper.pdf | |||||
2019 | Xidian University | A Survey of Deep Learning-based Object Detection | Object Detection | Salient object detection, pose detection | 網羅的なサーベイ | https://arxiv.org/pdf/1907.09408.pdf | |||||
2018 | City University of Hong Kong | Progressively Complementarity-aware Fusion Network for RGB-D Salient Object Detection | Object Detection | complementarity-aware fusion | 下記からの引用 | http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_Progressively_Complementarity-Aware_Fusion_CVPR_2018_paper.pdf | |||||
2019 | University of Western Australia | RGB-D image-based Object Detection: from Traditional Methods to Deep Learning Techniques | Object Detection | Deep fusion | fusion系検出のサーベイとして有用 | https://arxiv.org/pdf/1907.09236.pdf | |||||
2019 | Beijing University of Posts and Telecommunications, Columbia University | MemNet: Memory-Efficiency Guided Neural Architecture Search with Augment-Trim learning | NAS | MemNet | 従来はメモリが手動設計より4倍ほど大きかったのを従来程度にまで低減 | 56GPU days | ENASNet | https://arxiv.org/pdf/1907.09569.pdf | |||
2019 | Beijing Institute of Technology, University of Technology Sydney | Efficient Novelty-Driven Neural Architecture Search | NAS | EN2AS(efficient novelty-driven neural architecture search), a common search space used by Real et al. 2019 | 0.3GPU days | Random Search WS, ENAS | https://arxiv.org/pdf/1907.09109.pdf | ||||
2019 | Jadavpur University | Understanding Deep Learning Techniques for Image Segmentation | Semantic Segmentation | Unsupervised, Attention Models | 解釈、サーベイとして有用 | https://arxiv.org/pdf/1907.06119.pdf | |||||
1993 | 高速道路走行画像からの消失点推定に基づく実時間白線検出 | Lane Detection | https://www.jstage.jst.go.jp/article/ieejeiss1987/113/2/113_2_139/_pdf/-char/ja | ||||||||
2019 | ETH | A General Framework for Uncertainty Estimation in Deep Learning | Autonomous driving | adversarial attacks | https://arxiv.org/pdf/1907.06890.pdf | https://github.com/mattiasegu/A_General_Framework_for_Uncertainty_Estimation_in_Deep_Learning | |||||
2019 | Michigan State University | M3D-RPN: Monocular 3D Region Proposal Network for Object Detection | Road Marking RecognitionD Object Detection | BEV | Mono3D, 3DOP | https://arxiv.org/pdf/1907.06038.pdf | http://cvlab.cse.msu.edu/project-m3d-rpn.html | ||||
2019 | University of Waterloo, Facebook | Efficient Segmentation: Learning Downsampling Near Semantic Boundaries | Semantic Segmentation | content-adaptive downsampling | 不等間隔スパースサンプリングの補間により、速度と精度のバランスをとった。 | MioU>0.85(Supervisely), 演算量5BFlops以上は頭打ち | |||||
2018 | Intel | Apprentice: Using Knowledge Distillation Techniques To Improve Low-Precision Network Accuracy | Model Compression | 低ビット+distillation | https://openreview.net/pdf?id=B1ae1lZRb | ||||||
2019 | Peking University, Microsoft Research | Light Multi-segment Activation for Model Compression | Model Compression | -2.3%, 13.8%(CIFAR-10) | https://arxiv.org/pdf/1907.06870.pdf | ||||||
2018 | ETH, Google DeepMind | Model compression via distillation and quantization | Model Compression, Distillation | 上記からの引用 | Quantized Distillation | https://arxiv.org/pdf/1802.05668.pdf | https://github.com/antspy/quantized_distillation | ||||
2019 | ? | Seeing Through Fog Without Seeing Fog: Deep Sensor Fusion in the Absence of Labeled Training Data | Object Detection | SSD, pix2pix | LiDAR,RGB2系統のSSDを用意しMiddle Fusion。 入力層でdrop-out。 K-meansで解析し、層ごとのanchor boxを最適化。 | https://arxiv.org/pdf/1902.08913.pdfh | |||||
2019 | Horizon Robotics | VarGNet: Variable Group Convolutional Neural Network for Efficient Embedded Computing | NAS | Variable Group Convolutional Network (VarGNet) | Object Detection, Depthも評価, VGCでexpressiveness増加 | https://arxiv.org/pdf/1907.05653.pdf | |||||
2019 | Shanghai Jiao Tong University, Huawei | PC-DARTS: Partial Channel Connections for Memory-Efficient Differentiable Architecture Search | NAS | PC-DARTS | P-DARTSと大差なし | 0.1GPU-days | P-DARTS | https://arxiv.org/pdf/1907.05737.pdf | |||
2019 | University of Chinese Academy of Sciences | Cascade RetinaNet: Maintaining Consistency for Single-Stage Object Detection | Object Detection | Cas-RetinaNet, cascade, consistency | headを2つに分岐、consistencyを改善 | AP:41.1(COCO), 10fps(RTX 2080TI) | CornerNet511, Mask R-CNN | https://arxiv.org/pdf/1907.06881.pdf | |||
2017 | Tongji University | Scale Recovery for Monocular Visual Odometry Using Depth Estimated with Deep Convolutional Neural Fields | Visual odometry | maximum likelihood | ResNetでデプス推定,VO・depthをCRFで統合 | T:2.68%,R: 0.0022deg./m(KITTI) | VISO2 | http://openaccess.thecvf.com/content_ICCV_2017/papers/Yin_Scale_Recovery_for_ICCV_2017_paper.pdf | |||
2019 | University of T¨ubingen | Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming | Object Detection | Style transfer as data augmentation | 霧・雪・ノイズ・ブラー耐性をstyle transfer DAで向上。モデル・データセットの比較が充実。 | https://arxiv.org/pdf/1907.07484.pdf | https://github.com/bethgelab/robust-detection-benchmark | ||||
2017 | Road detection based on the fusion of Lidar and image data | Free-space | LiDAR, fusion, Adaboost | https://journals.sagepub.com/doi/pdf/10.1177/1729881417738102 | |||||||
2019 | UESTC, ETH, Google | DeepLiDAR: Deep Surface Normal Guided Depth Prediction for Outdoor Scene from Sparse LiDAR Data and Single Color Image | Depth | LiDAR | 法線推定 KW:lidar fusion road surface geometry | http://openaccess.thecvf.com/content_CVPR_2019/papers/Qiu_DeepLiDAR_Deep_Surface_Normal_Guided_Depth_Prediction_for_Outdoor_Scene_CVPR_2019_paper.pdf | |||||
2019 | Shanghai Jiao Tong University, Huawei | PC-DARTS Partial Channel Connections for Memory-Efficient Differentiable Architecture Search | NAS | sampling a small part of super-net to reduce the redundancy | Fast Houghの詳細が不明、影には強そう | 0.1GPU days, 2.57%(CIFAR10) | ProxylessNAS(4GPU daysだが精度は最高) | https://arxiv.org/pdf/1907.05737.pdf | https://github.com/yuhuixu1993/PC-DARTS | ||
2019 | Sensor Cortek Inc. | How much real data do we actually need: Analyzing object detection performance using synthetic and real data | Object Detection | driving simulator | SSD+MobileNetで、学習画像の何%までシミュレータ画像に置き換えられるか | https://arxiv.org/pdf/1907.07061.pdf | |||||
2019 | Michigan State University | M3D-RPN: Monocular 3D Region Proposal Network for Object Detection | https://arxiv.org/pdf/1907.06038.pdf | http://cvlab.cse.msu.edu/project-m3d-rpn.html. | |||||||
2018 | Target Fusion Detection of LiDAR and Camera Based on the Improved YOLO Algorithm | Object Detection | YOLO, LiDAR | LiDARをデプス画像に変換してearly fusion, 2出力出してまたfusion | https://www.mdpi.com/2227-7390/6/10/213/pdf | ||||||
2019 | University of Michigan, Baidu | Adversarial Objects Against LiDAR-Based Autonomous Driving Systems | Object Detection | LiDAR-Adv, adversarial example, Baidu Apollo autonomous driving platform | https://arxiv.org/pdf/1907.05418.pdf | https://sites.google.com/view/lidar-adv | |||||
1999 | Universita di Verona | A Compact Algorithm for Rectification of Stereo Pairs | Rectification | stereo rectification | MATLAB22行で実装 | http://www.diegm.uniud.it/fusiello/papers/mva99.pdf | |||||
2019 | Google, Baidu | EPNAS: Efficient Progressive Neural Architecture Search | NAS | REINFORCE, performance prediction, scale LSTM+insert LSTM, more efficient than perlayer modification of ENAS | DARTSと同速度, FLOPs/byte,MFLOPsで評価 | DenseNet, MNASNet | https://arxiv.org/pdf/1907.04648.pdf | ||||
2019 | Tohoku University | Joint Learning of Multiple Image Restoration Tasks | Image restoration | single input and multiple output, attention mechanism, deblur, dehaze | SEの変形をattentionと称している | https://arxiv.org/pdf/1907.04508.pdf | |||||
2019 | University of Washington | Sparse Networks from Scratch: Faster Training without Losing Performance | Model Compression | sparse momentum | -6.2% w/ 10% weights(ImageNet) | https://arxiv.org/pdf/1907.04840.pdf | |||||
2017 | 中央大学 | 正距円筒画像のステレオ平行化 | Stereo Rectification | 魚眼ステレオカメラ | 対応点探索はテンプレートマッチング(SAD), 内部パラメータはMatlab の OcamCalib Toolboxを利用 | http://www.mech.chuo-u.ac.jp/umedalab/publications/pdf/2016/201703_ohashi_DIA.pdf | |||||
2008 | ETH Zurich | Automatic Detection of Checkerboards on Blurred and Distorted Images | Camera Calibration | Adaptation of Erosion Kernels | 魚眼向け、ボケ・歪みにロバストなチェッカーボード検出方法、上記からの引用 | Vezhnevets | http://rpg.ifi.uzh.ch/docs/IROS08_scaramuzza_b.pdf | ||||
2019 | Facebook AI Research | Panoptic Feature Pyramid Networks | Panoptic Segmentation | Mask R-CNN, FPN | Mask R-CNNの改良 | PQ_Th46.8(COCO), 速度不明 | DeeplabV3+ | https://arxiv.org/pdf/1901.02446.pdf | |||
2019 | Uber | UPSNet: A Unified Panoptic Segmentation Network | Panoptic Segmentation | FPN+semantic/instance head+panoptic head, セマセグで路面標示は認識しない, 輪郭がなまっている | PQ46.6, 171ms(COCO, GTX 1080 Ti) | MR-CNN-PSP, AUNet, Megvii (Face++) | https://arxiv.org/pdf/1901.03784.pdf | https://github.com/uber-research/UPSNet | https://www.slideshare.net/ShunsukeNakamura17/cvpr-2019-report-30-papers/ | ||
2019 | University of Chinese Academy of Sciences | C-MIL Continuation Multiple Instance Learning for Weakly Supervised Object Detection | Object Detection | Weakly supervised object detection, continuation multiple instance learning (C-MIL) | 弱教師あり(クラスのみ)、Lossを段階的に難しく(詳細に)していく | mAP53.1(VOC 2007) | OICR, MELM | https://arxiv.org/pdf/1904.05647.pdf | |||
2019 | ETRI | An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection | Object Detection | VoVNet | Real Time/Pascal2007でSOTA(Browse state-of-the-art) | 16.6FPS, AP38.5 | CornerNet | https://arxiv.org/pdf/1904.09730v1.pdf | https://paperswithcode.com/sota/real-time-object-detection-on-pascal-voc-2007 | ||
2019 | Amazon Web Services | A Unified Optimization Approach for CNN Model Inference on Integrated GPUs | Acceleration | integrated GPUs at the edge | 1.23~39.3x | OpenVINO on AWS DeepLens, ACL on Acer aiSage | https://arxiv.org/pdf/1907.02154.pdf | ||||
2019 | NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection | Object Detection, NAS | RNNController | +~2% than FPN | FPN | https://arxiv.org/pdf/1904.07392.pdf | https://www.slideshare.net/ShunsukeNakamura17/cvpr-2019-report-30-papers/ https://qiita.com/takoroy/items/b081c7e7fa9fe6b11abd https://www.slideshare.net/DeepLearningJP2016/dlnasfpn-learning-scalable-feature-pyramid-architecture-for-object-detection-145224193 | ||||
2019 | University of Pisa | A Survey Of Methods For Explaining Black Box Models | Interpretability | Single Tree Approximation, Rule Extraction, Agnostic Explanator | 解釈系のサーベイなので重要 | https://arxiv.org/pdf/1802.01933.pdf | |||||
2018 | Tsinghua University, SenseTime | Quantization Mimic: Towards Very Tiny CNN for Object Detection | Model Compression | uniform quantization, mimic learning, R-FCN, Faster R-CNN | 物体認識が対象なのが珍しい | Acc. +1.1%, 604x(size)/10.8x(speed)(VGG, WIDER FACE, medium) | https://arxiv.org/pdf/1805.02152.pdf | ||||
2019 | Model Compression by Entropy Penalized Reparameterization | Model Compression | entropy penalty, scalar quantization, DFT | priorに対する向上は小さい | Acc. -3.4%, 590x(VGG-16 on CIFAR-10) | Minimal Random Code Learning | https://arxiv.org/pdf/1906.06624.pdf | ||||
2018 | University of Cambridge | Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters | Model Compression | random sample, Kullback-Leibler divergence | 上記Googleからの引用 | Acc. -0.07%, 159x(VGG-16 on CIFAR-10) | Bayesian Compression | https://arxiv.org/pdf/1810.00440.pdf | https://github.com/cambridge-mlg/miracle (Tensorflow) | ||
2019 | Northeastern University | AutoSlim: An Automatic DNN Structured Pruning Framework for Ultra-High Compression Rates | Model Compression | ADMM-based structured weight pruning | Acc. -0%, 9.2x(ResNet-50 using ImageNet) | AMC | https://arxiv.org/pdf/1907.03141.pdf | ||||
2019 | University of Bologna | Lane Detection and Classification using Cascaded CNNs | Lane detection | 58.93fps, Accuracy95.24(TuSimple) | https://arxiv.org/pdf/1907.01294.pdf | https://github.com/fabvio/Cascade-LD | |||||
2019 | Universidade Federal do Espírito Santo | Traffic Light Recognition Using Deep Learning and Prior Maps for Autonomous Cars | TLR | YOLOv3 | ELASのLab.LCAD | Precision80.86(DTLD) | https://arxiv.org/pdf/1906.11886.pdf | https://github.com/LCAD-UFES/carmen_lcad/blob/master/src/traffic_light_yolo/README.md | |||
上へ移動 | |||||||||||
2019 | Cornell University | Confidence Calibration for Convolutional Neural Networks Using Structured Dropout | Training technique | Structured Dropout | https://arxiv.org/pdf/1906.09551.pdf | ||||||
2019 | MIT | Placeto: Learning Generalizable Device Placement Algorithms for Distributed Machine Learning | device | Placeto, RL, distributed neural network training | https://arxiv.org/pdf/1906.08879.pdf | ||||||
2019 | Huazhong University of Science and Technology, Horizon Robotics | Densely Connected Search Space for More Flexible Neural Architecture Search | NAS | differentiable, densely connected search space | 3.8GPU days | DARTS, ... | https://arxiv.org/pdf/1906.09607.pdf | ||||
2019 | Goodyear | Deep Learning in the Automotive Industry:Recent Advances and Application Examples | Automotive | 車載の概観が有用だが内容が薄め | https://arxiv.org/ftp/arxiv/papers/1906/1906.08834.pdf | ||||||
2019 | ?(Tobias Gruber) | Pixel-Accurate Depth Evaluation in Realistic Driving Scenarios | Depth Benchmark | 種々のモダリティ・天候の比較が貴重 車載向け | RGB+Lidar 2017Unsupervised monocular depth estimation with left-right consistency | https://arxiv.org/pdf/1906.08953.pdf | |||||
2019 | University of California, San Diego | Cascade R-CNN: High Quality Object Detection and Instance Segmentation | Object Detection | Cascade R-CNN, trained with increasing IoU | https://arxiv.org/pdf/1906.09756.pdf | https://github.com/zhaoweicai/cascade-rcnn (Caffe) https://github.com/zhaoweicai/Detectron-Cascade-RCNN (Detectron) | |||||
2019 | Alibaba | XNAS: Neural Architecture Search with Expert Advice | NAS | 0.3GPU days | DARTS | https://arxiv.org/pdf/1906.08031.pdf | |||||
2019 | ETH, Google | Transfer NAS: Knowledge Transfer between Search Spaces with Transformer Agents | NAS | Attention-based agent | Random Search | https://arxiv.org/pdf/1906.08102.pdf | |||||
2019 | Duke University | SwiftNet: Using Graph Propagation as Meta-knowledge to Search Highly Representative Neural Architectures | NAS | SwiftNet, GRAph propagation as Meta-knowledge | 0.35GPU days | DARTS, ... | https://arxiv.org/pdf/1906.08305.pdf | ||||
2019 | A real-time road detection method based on reorganized lidar data | Road Boundary Detection | LiDAR併用 | Fast Houghの詳細が不明、影には強そう | , AP:81.98~89.97% | https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0215159&type=printable | |||||
2019 | Deep Learning for Large-Scale Traffic-Sign Detection and Recognition | TSR | Mask R-CNN(mask使ってないので、ただのROI-AlignつきFasterか)、Detectron利用 | , | https://arxiv.org/pdf/1904.00649.pdf | ||||||
2019 | Driver Behavior Analysis Using Lane Departure Detection Under Challenging Conditions | Lane Detection | Mask R-CNN | 自車線のfree spaceしか出ない | 25fps(752x480), | https://arxiv.org/pdf/1906.00093.pdf | |||||
2019 | Enhanced free space detection in multiple lanes based on single CNN with scene identification | Road Boundary Detection | CNN(16GMAC) | 22.59fps(Titan Xp, VGA), | https://arxiv.org/pdf/1905.00941.pdf | https://github.com/fabvio/ld-lsi/. | |||||
2019 | Improving benchmarks for autonomous vehicles testing using synthetically generated images | TSR | DCGAN | , | https://arxiv.org/pdf/1904.10261.pdf | ||||||
2019 | Improving Traffic Sign Detection by Combining MSER and Lucas Kanade Tracking | TSR | 47.95fps, 0.9353 | http://www.ijicic.org/ijicic-150216.pdf | |||||||
2019 | Real-Time Embedded Traffic Sign Recognition Using Efficient Convolutional Neural Network | TSR | Det:EmdNet(6.3M)/cls:ENet(0.9M) | 33ms(GTX2080,400x260)/0.62ms(32x32), 98.6% | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8698449 | ||||||
2019 | Robust multi-lane detection and tracking using adaptive threshold and lane classification | Lane Detection | 動画あり、あまり遠くまで検出していないが、安定している | , | https://link.springer.com/article/10.1007/s00138-018-0977-0 | ||||||
2019 | The Right (Angled) Perspective: Improving the Understanding of Road Scenes using Boosted Inverse Perspective Mapping | Road Marking Recognition | GANを使ったIPM | , | https://arxiv.org/pdf/1812.00913.pdf | ||||||
2019 | Traffic Light Recognition Using Deep Learning and Prior Maps for Autonomous Cars | TLR | YOLOv3, uses Darknet-53 | 47ms(608×608,TITAN), 55.21% of mAP on IARA-TLD | https://arxiv.org/pdf/1906.11886.pdf | https://github.com/LCAD-UFES/carmen_lcad/blob/master/src/traffic_light_yolo/README.md | |||||
2019 | Nagoya University | Traffic light recognition using high-definition map features | TLR | SSD(Caffe), ROI(LiDAR前提,高速化の効果は微小) | Daytime/Sunsetのみ評価 | 18ms(980 Ti), | https://reader.elsevier.com/reader/sd/pii/S0921889018301234?token=C3AF760DE57F8F007FEF0FFA7CBC4FFE1A6F573B8B3D07991602198878A76F01CBE86023B86697364BF02BA27FD2B110 | ||||
2019 | Traffic Sign Classification Using Computationally Efficient Convolutional Neural Networks | TSR | CNN | 失敗パターンの詳細な分析、MAC評価 | , | http://www.diva-portal.org/smash/get/diva2:1324051/FULLTEXT01.pdf | |||||
2019 | National Chung Hsing University | Ultra-Low Complexity Block-Based Lane Detection and Departure Warning System | Lane Detection | 高速、いろいろ有用な概念 | 4.28ms(FHD), departure warning rate are 96.12% | http://www.ee.nchu.edu.tw/Pic/Writings/3985_publish-version.pdf | |||||
2019 | ? | Vehicle Detection and Speed Estimation for Automated Traffic Surveillance Systems at Nighttime | Lane Detection,Head Light Detection,Tail Light Detection | 単純なrule-based | , PSNR | https://hrcak.srce.hr/file/316826 | |||||
2018 | Chongqing University | A Fast Learning Method for Accurate and Robust Lane Detection Using Two-Stage Feature Extraction with YOLO v3 | Road Boundary Detection | CNN(YOLOv3) | , mAP:88.39% | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6308794/pdf/sensors-18-04308.pdf | |||||
2018 | University Of Waterloo | A Hierarchical Deep Architecture and Mini-Batch Selection Method For Joint Traffic Sign and Light Detection | TSR,TLR | CNN(Faster R-CNN with ResNet-50→crop→分類) | 速い分精度が低い SOTAとされている | 15ms(1080Ti), Acc:0.68 Boschのベンチでは~150ms | https://arxiv.org/pdf/1806.07987.pdf | https://paperswithcode.com/task/traffic-sign-recognition/codeless | https://github.com/bosch-ros-pkg/bstld | ||
2018 | Government Women Residence Polytechnic | A review of recent advances in lane detection and departure warning system | Road Boundary Detection | , | https://www.sciencedirect.com/science/article/abs/pii/S0031320317303266 | ||||||
2018 | Changchun University of Science and Technology | A Robust Lane Detection Method Based on Vanishing Point Estimation | Lane Detection,Road Boundary Detection | line segment detector (LSD) | Houghの計算量の課題への対策 | , Correct detection rate:92.34%~ | https://ac.els-cdn.com/S1877050918305489/1-s2.0-S1877050918305489-main.pdf?_tid=fc3318c9-ce4d-498f-bd80-49b939d8b568&acdnat=1552210218_7f1ec4dd45fdd6611c3c2ff1aad1ff51 | ||||
2018 | Hosei University | A Study on Fast and Robust Vanishing Point Detection System Using Fast M-Estimation Method and Regional Division for In-vehicle Camera | M-Estimation | 消失点推定のみ、A Robust Lane Detection Method Based on Vanishing Point Estimationと組み合わせられそう | , | http://www.davidpublisher.org/Public/uploads/Contribute/5a98f6da0975e.pdf | |||||
2018 | Huazhong University of Science and Technology | Cascaded segmentation-detection networks for text-based traffic sign detection | TSR | テキストベースなので除外 | , | https://ieeexplore.ieee.org/document/8239744 | |||||
2018 | Beihang University | Deep detection network for real-life traffic sign in vehicular networks | TSR | CNN(Attention Network+Faster-RCNN) | , mAP of 80.31% and 94.95% in two benchmarks | https://www.sciencedirect.com/science/article/abs/pii/S1389128618300999 | |||||
2018 | Universidad de Sevilla | Deep neural network for traffic sign recognition systems: An analysis of spatial transformers and stochastic optimisation methods | TSR | CNN(simple+spatial transformer modules) | , 99.71% | https://idus.us.es/xmlui/bitstream/handle/11441/80679/NEUNET-D-17-00381.pdf?sequence=1 | |||||
2018 | FZI Research Center for Information Technology | Deep Semantic Lane Segmentation for Mapless Driving | Road Boundary Detection | CNN | Lidar CNN [14]は94.07%, 先行例に大幅に負けているが...、KITTI10位以内 | , F1:88.89% | https://www.mrt.kit.edu/z/publ/download/2018/Meyer2018SemanticLanes.pdf | ||||
2018 | UC Berkeley | Deep Traffic Light Detection for Self-driving Cars from a Large-scale Dataset | TLR | SSD | Fusion転用にしても後段の演算コスト高くないか? | real-timeとだけ記述(288x512にリサイズ,Titan X), | https://ywpkwon.github.io/pdf/18itsc.pdf | ||||
2018 | Ulm University | Detecting Traffic Lights by Single Shot Detection | TLR | SSD | データを増やすほど精度が向上した。230K。公開コード 物体が小さいほど context情報が重要 | 101ms, log-average miss rate<0.02(DriveU) | https://arxiv.org/pdf/1805.02523.pdf | https://github.com/julimueller/tl_ssd | |||
2018 | Efficient Road Lane Marking Detection with Deep Learning | Lane Detection | 29.1ms(1080), | https://arxiv.org/ftp/arxiv/papers/1809/1809.03994.pdf | |||||||
2018 | Efficient Traffic-Sign Recognition with Scale-aware CNN | TSR | CNN、Pyramid入力→RP:backbone途中から2分岐convx4→cls:Inception Module+FC、Online Hard Example Mining | 226ms(RP)+56ms(cls), 99.88% precision | https://arxiv.org/pdf/1805.12289.pdf | ||||||
2018 | Universidade Federal do Espエirito Santo | Ego-Lane Analysis System (ELAS) Dataset and Algorithms | Lane Detection,Road Boundary Detection,Road Marking Recognition | 路面標示は32にリサイズしてNCCでテンプレートマッチング | , MAEのみ評価 | ||||||
2018 | Northwestern Polytechnical University | Embedding Structured Contour and Location Prior in Siamesed Fully Convolutional Networks for Road Detection | Road Boundary Detection | , F1:95.36 | https://ieeexplore.ieee.org/document/8058005 | ||||||
2018 | Universidad de Sevilla | Evaluation of deep neural networks for traffic sign detection systems | TSR | CNN(Faster,SSD,YOLO) | COCOでpre-train/GTSDBでfine-tune | , | https://www.sciencedirect.com/science/article/pii/S092523121830924X?via%3Dihub | https://github.com/aarcosg/traffic-sign-detection | |||
2018 | University of Engineering Science | Lane Detection Based on Connection of Various Feature Extraction Methods | Road Boundary Detection | Hough,Canny+Kalman filter | , | https://www.hindawi.com/journals/am/2018/8320207/ | |||||
2018 | LIDAR-Camera Fusion for Road Detection Using Fully Convolutional Neural Networks | Road Boundary Detection | LiDAR fusion | , | https://arxiv.org/pdf/1809.07941.pdf | ||||||
2018 | University of Shanghai | Localized Traffic Sign Detection with Multi-scale Deconvolution Networks | TSR | CNN(Modified Feature Pyramid Network) | Chinese Traffic Sign Dataset | , <99.1% | https://arxiv.org/pdf/1804.10428 | ||||
2018 | Mark Yourself Road Marking Segmentation via Weakly-Supervised Annotations from Multimodal Data | Road Marking Recognition | CRF | 路面標示のセマセグアノテーション生成 | , | http://www.robots.ox.ac.uk/~mobile/Papers/2018ICRA_bruls.pdf | |||||
2018 | University of Waterloo | MicronNet A Highly Compact Deep Convolutional Neural Network Architecture for Real-time Embedded Traffic Sign Classification | TSR | CNN(0.51M) | 分類のみ、軽量、NAS?、全画像48x48にリサイズしてから評価 | 32.19ms(Cortex-A53), 98.9% | https://arxiv.org/pdf/1804.00497.pdf | ||||
2018 | Real-Time Detection and Recognition of Road Traffic Signs using MSER and Random Forests | TSR | SVM→RFで10%程度向上、CNN(上記2016Traffic sign...)と比較してわずかに低い | 469ms, 88.83~92.67% | |||||||
2018 | University of Banja Luka | Real-time Large Scale Traffic Sign Detection | TSR | CNN(YOLOv3) | , mAP>88% | https://ieeexplore.ieee.org/abstract/document/8587013 | |||||
2018 | University of Science and Technology of China | Real-Time Traffic Sign Recognition Based on Efficient CNNs in the Wild | TSR | CNN(Faster-RCNN) | , ? | https://ieeexplore.ieee.org/abstract/document/8392744 | |||||
2018 | Normandie Univ | Recent Advances in Object Detection in the Age of Deep Convolutional Neural Networks | , | https://arxiv.org/pdf/1809.03193.pdf | |||||||
2018 | Hirosaki University | Road Boundary Detection using In-vehicle Monocular Camera | Road Boundary Detection | , | https://www.scitepress.org/papers/2018/65897/65897.pdf | ||||||
2018 | Worcester Polytechnic Institute | Road Segmentation Using CNN with GRU | Road Boundary Detection | gated recurrent units | 高速だが従来DNNより精度が若干劣る。歩道にFP。2016Exploitingと大差なさそう。 | 20ms(GTX 950M), F1:86.91 %,AP:81.11%(KITTI40位くらいのレベル) | https://arxiv.org/pdf/1804.05164.pdf | ||||
2018 | Robust Free-Space Detection in Urban Roads based on MSER Extraction Using Gradient Images | Road Boundary Detection | ピックアップ済み。 | 18.34s, | https://ieeexplore.ieee.org/document/8482557 | ||||||
2018 | Kyungpook National University | Robust Lane Detection Algorithm Based on Triangular Lane Model | Lane Detection | 曲線に対応、近・遠分けて推定 | 記載なし, | http://www.pmf.ni.ac.rs/filomat-content/2018/32-5/32-5-15-6663.pdf | |||||
2018 | Robust Lane Detection for Complicated Road Environment Based on Normal Map | Lane Detection,Road Boundary Detection | デプス(ステレオ前提)から法線を推定 | , | https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8456509 | ||||||
2018 | Qualcomm | Simultaneous traffic sign detection and boundary estimation using convolutional neural network | TSR | CNN | , | https://arxiv.org/pdf/1802.10019 | |||||
2018 | The Architectural Implications of Autonomous Driving: Constraints and Acceleration | TSR | YOLO | 車メイン | , | http://web.eecs.umich.edu/~yunqi/pdf/lin2018autonomous.pdf | |||||
2018 | Total Recall Understanding Traffic Signs using Deep Hierarchical Convolutional Neural Networks | TSR | CNN(6.3M) | 複雑なskipとdilatedを組み合わせた新しい構造 | , 99.33%/99.17% | https://arxiv.org/pdf/1808.10524.pdf | |||||
2018 | ESAT-PSI | Towards End-to-End Lane Detection: an Instance Segmentation Approach | Lane Detection | CNN(seg+fitting), conditional homography using H-Net, Pixel embeddings+Binary lane segmentationの2系統のANDを取りクラスタリング | TuSimple Benchmark 3位、実装難しそう、FPあったりカーブが明らかに大外れしてたり気になる点も。 | 19ms(1080Ti), 96.4%, MSE:5.99[pix] | https://arxiv.org/pdf/1802.05591.pdf | ||||
2018 | Tsinghua University | Traffic light recognition for complex scene with fusion detections | TLR | ACF, fuzzy detection | 精度がやたら低いのが気になる。難シーン?bulbと全体を別に検出 | 81ms(CPU), | https://wangxiang10.github.io/papers/ITS18.pdf | ||||
2018 | Traffic Sign Detection based on Color Segmentation of Obscure Image Candidates: A Comprehensive Study | TSR | サーベイ | , | http://www.mecs-press.org/ijmecs/ijmecs-v10-n6/IJMECS-V10-N6-5.pdf | ||||||
2018 | Utrecht University | Traffic sign recognition using a multi-task convolutional neural network | TSR | CNN(Modified Feature Pyramid Network) | , | http://ir.ia.ac.cn/bitstream/173211/19824/1/TITS-2017-HengliangLuo.pdf | |||||
2018 | Traffic Signs Recognition and Classification based on Deep Feature Learning | TSR | , | ||||||||
2018 | Tutorial: Build a lane detector | Road Boundary Detection | 論文ではない。Hough変換ベースの基本的な手法のPythonコードによる解説。 | , | https://towardsdatascience.com/tutorial-build-a-lane-detector-679fd8953132 | ||||||
2018 | The Hong Kong Polytechnic University | Vehicle detection in intelligent transportation systems and its applications under varying environments: A review | Tail Light Detection | テールランプはほとんど記述なし | , | https://www.researchgate.net/profile/Zi_Yang12/publication/320370175_Vehicle_Detection_in_Intelligent_Transportation_Systems_and_its_Applications_under_Varying_Environments_A_Review/links/5ad55e75458515c60f546f95/Vehicle-Detection-in-Intelligent-Transportation-Systems-and-its-Applications-under-Varying-Environments-A-Review.pdf | |||||
2018 | リコー | 車載カメラにおける信号機認識および危険運転イベント検知 | TLR | YUV色特徴 | , | https://jp.ricoh.com/technology/techreport/43/pdf/RTR43a02.pdf | |||||
2017 | 輝度勾配方向を用いたエッジのクラスタリングによる一般道の白線検出 | Lane Detection | LSD | , | https://www.jstage.jst.go.jp/article/transjsme/83/854/83_17-00095/_pdf | ||||||
2017 | Samara National Research University | A Method for Traffic Sign Recognition with CNN using GPU | TSR | full HDで50m先の標識まで認識可能 | 50fps(64x64), 99.94% | http://www.scitepress.org/Papers/2017/64361/64361.pdf | |||||
2017 | Changsha University of Science and Technology | A Real-Time Chinese Traffic Sign Detection Algorithm Based on Modified YOLOv2 | TSR | det:GHT, cls:CNN | , | https://www.researchgate.net/publication/321117701_A_Real-Time_Chinese_Traffic_Sign_Detection_Algorithm_Based_on_Modified_YOLOv2 | |||||
2017 | Northwestern Polytechnical University | An Incremental Framework for Video-Based Traffic Sign Detection, Tracking, and Recognition | TSR | , | https://crabwq.github.io/pdf/2017%20An%20Incremental%20Framework%20for%20Video-Based%20TrafficSign%20Detection,%20Tracking,%20and%20Recognition.pdf | ||||||
2017 | University of Batna 2 | An overview of traffic sign detection and classification methods | TSR | DL/non DLの一覧がある | 以下全てこの引用 | , | https://link.springer.com/article/10.1007/s13735-017-0129-8 | ||||
2017 | An Overview of Traffic Signs Recognition Methods | TSR | , | https://pdfs.semanticscholar.org/5fd7/eca5432e4e4038dfc2ad425fa5f3e8739f0b.pdf | |||||||
2017 | University of Bedfordshire | Car make and model recognition under limited lighting conditions at night | Tail Light Detection | genetic algorithm | 車種分類が主眼 | , | https://link.springer.com/content/pdf/10.1007%2Fs10044-016-0559-6.pdf | ||||
2017 | CNN Design for Real-Time Traffic Sign Recognition | TSR | , | https://kundoc.com/pdf-cnn-design-for-real-time-traffic-sign-recognition-.html | |||||||
2017 | Deep learning traffic sign detection, recognition and augmentation | TSR | , | https://www.researchgate.net/publication/317607639_Deep_learning_traffic_sign_detection_recognition_and_augmentation | |||||||
2017 | University of South Carolina | Detecting Small Signs from Large Images | TSR | 標識サイズ分類 | , | https://www.researchgate.net/publication/317954746_Detecting_Small_Signs_from_Large_Images | |||||
2017 | Embedding vision-based advanced driver assistance systems a survey | Road Boundary Detection,TSR,Tail Light Detection | 速度に特化したサーベイ | , | https://www.researchgate.net/publication/307544971_Embedding_Vision-based_Advanced_Driver_Assistance_Systems_a_Survey | ||||||
2017 | Nehru Institute of Technology | Enhanced Technique for Nighttime Vehicle Detection with Multiple Features | Tail Light Detection | 記載なし, | https://www.onlinejournal.in/IJIRV3I1/279.pdf | ||||||
2017 | Inha University | Fast Lamp Pairing–based Vehicle Detection Robust to Atypical and Turn Signal Lamps at Night | Tail Light Detection | ~6ms, 90.9~96.1% | http://www.ieiespc.org/include/pdfdownload.asp?filename=IEEKSPC_2017_6_4_20170830143515_1.pdf&xc=318 | ||||||
2017 | Zhejiang University of Technology | Faster R-CNN for Small Traffic Sign Detection | TSR | objectサイズのヒストグラム 受容野が適切でないことが精度低下の原因 入力upsamplingでも精度向上 SSDでは収束しないし低精度 | , | https://link.springer.com/chapter/10.1007/978-981-10-7305-2_14 | |||||
2017 | Hybrid conditional random field based camera-LIDAR fusion for road detection | Road Boundary Detection | CRF | LiDAR fusion | , | http://www.massey.ac.nz/~rwang/publications/17-IS-Xiao.pdf | |||||
2017 | Near East University | Lane detection by estimating and using restricted search space in Hough domain | Road Boundary Detection | SymmetricalLocalThreshold (SLT)+polar Hough | , | https://ac.els-cdn.com/S1877050917324341/1-s2.0-S1877050917324341-main.pdf?_tid=a14bf3de-a5b2-4320-80d9-68c812a1ab44&acdnat=1552210211_a031bbc00c4a5e27ea8f36f02e0f785c | |||||
2017 | A.I.S.S.M.S. Institute of Information Technology | Night-time Vehicle Detection for Automatic Headlight Beam Control | Tail Light Detection | SVM | FPやたら多い | , 94.8453608247423 | https://pdfs.semanticscholar.org/694b/51965133a7bc3477b6c56e6a8322cff3c9ab.pdf | ||||
2017 | Vietnam National University | Nighttime vehicle detection and classification via headlights trajectories matching | Head Light Detection | , 81.19% | https://ieeexplore.ieee.org/abstract/document/8030869 | ||||||
2017 | City University of Hong Kong | Nighttime vehicle detection based on bio-inspired image enhancement and weighted score-level feature fusion | Tail Light Detection | 230ms, 94.6% | http://www.ee.cityu.edu.hk/~lhchan/Publications/2017%20IEEE%20TITS%20Kuang.pdf | ||||||
2017 | Partitioning of the Free Space-Time for On-Road Navigation of Autonomous Ground Vehicles | Road Boundary Detection | , | https://hal-mines-paristech.archives-ouvertes.fr/hal-01691793/document | |||||||
2017 | Beijing Institute of Technology | Perceptual Generative Adversarial Networks for Small Object Detection | TSR | GAN, small object | 600ms, Det:88%(Tsinghua-Tencent 100K) | http://openaccess.thecvf.com/content_cvpr_2017/papers/Li_Perceptual_Generative_Adversarial_CVPR_2017_paper.pdf | |||||
2017 | Real-time traffic sign recognition based on a general purpose GPU and deep-learning | TSR | ACFと比べ精度格段に向上 | , | https://www.researchgate.net/publication/314269407_Real-time_traffic_sign_recognition_based_on_a_general_purpose_GPU_and_deep-learning | ||||||
2017 | Road Lane Detection Robust to Shadows Based on a Fuzzy System Using a Visible Light Camera Sensor | Lane Detection,Road Boundary Detection | ピックアップ済み。手法間のPro/Con表が参考になる。ラインの輪郭を抽出。 | 24.77ms, Precision:0.78~0.97 | https://www.researchgate.net/publication/320724487_Road_Lane_Detection_Robust_to_Shadows_Based_on_a_Fuzzy_System_Using_a_Visible_Light_Camera_Sensor/download | ||||||
2017 | MediaTek Inc. | Robust detection and tracking of vehicle taillight signals using frequency domain feature based adaboost learning | Turn Light Detection | ref.7個、被引用1件、価値低そう | , | https://ieeexplore.ieee.org/abstract/document/7991176/ | |||||
2017 | KAIST | Robust Road Marking Detection and Recognition Using Density-Based Grouping and Machine Learning Techniques | Road Marking Recognition | PCANet + Logistic Regression | 10class | 記載なし, 82.1~100% | https://www.researchgate.net/profile/Oleksandr_Bailo/publication/312286941_Robust_Road_Marking_Detection_and_Recognition_Using_Density-Based_Grouping_and_Machine_Learning_Techniques/links/59f0228f0f7e9baeb26ad615/Robust-Road-Marking-Detection-and-Recognition-Using-Density-Based-Grouping-and-Machine-Learning-Techniques.pdf | ||||
2017 | Symbolic Road Marking Recognition Using Convolutional Neural Networks | Road Marking Recognition | LeNet | 10class、14479パッチ出学習、評価がざっくり。 | 記載なし, 99.05% | https://www.cse.unr.edu/~bebis/IV17.pdf | |||||
2017 | University of Ljubljana | Towards large-scale traffic sign detection and recognition | TSR | CNN(Faster R-CNN) | FNの82%は50画素以下 | , mAP:All 123 classes 79.41 | http://prints.vicos.si/publications/files/351 | ||||
2017 | Traffic Light Recognition Exploiting Map and Localization at Every Stage | TLR | 矢印もスコープ(ただし韓国の特殊な4連タイプ) | , Precision:98.68% | https://www.researchgate.net/publication/318385033_Traffic_Light_Recognition_Exploiting_Map_and_Localization_at_Every_Stage | ||||||
2017 | Sun Yat-sen University | Turn signal detection during nighttime by CNN detector and perceptual hashing tracking | Turn Light Detection | , | https://ieeexplore.ieee.org/abstract/document/7891988/ | ||||||
2017 | Technical University of Cluj-Napoca | Vehicle taillight detection and tracking using deep learning and thresholding for candidate generation | Tail Light Detection | CNN | , | https://ieeexplore.ieee.org/abstract/document/8117015/ | |||||
2017 | KAIST,Samsung | VPGNet: Vanishing Point Guided Network for Lane and Road Marking Detection and Recognition | Lane Detection,Road Marking Recognition | CNN(8層) | 17class分類,Caffe実装公開 | ~50ms w/ GTX Titan X, F1:0.743~0.87 Lane detection score:0.5234 2016Traffic-sign detection and classification in the wildより1-2%いい | https://arxiv.org/abs/1710.06288 | https://github.com/SeokjuLee/VPGNet/blob/master/caffe/matlab/demo/classification_demo.m | |||
2017 | 奈良先端 | ステレオビジョンによる道路領域の推定 | Road Boundary Detection | , | https://library.naist.jp/mylimedio/dllimedio/showpdf2.cgi/DLPDFR013327_P1_50 | ||||||
2016 | University Rovira i Virgili | A practical approach for detection and classification of traffic signs using Convolutional Neural Networks | TSR | CNN | 37.72fps, 99.89%(detection),99.55%(classification) | https://www.sciencedirect.com/science/article/pii/S092188901530316X | |||||
2016 | National Chiao Tung University | An Image Based Overexposed Taillight Detection Method for Frontal Vehicle Detection in Night Vision | Tail Light Detection | 過剰露出の課題への対策 | 30fps, 98.6% | http://www.apsipa.org/proceedings_2016/HTML/paper2016/63.pdf | |||||
2016 | Institute for Infocomm Research Agency for Science, Technology and Research (A∗STAR) | Appearance-based brake-lights recognition using deep learning and vehicle detection | Tail Light Detection | 車両認識を同時に行う、LiDAR fusion | , | https://ieeexplore.ieee.org/abstract/document/7535481 | |||||
2016 | Department of EEE | Autonomous Tracking of Vehicle Taillights for Avoiding Accidents in Toll Road | Tail Light Detection | 引用[8-10,14]が参考になるかも | , | https://www.ijraset.com/fileserve.php?FID=5009 | |||||
2016 | National Chiao Tung University | Daytime preceding vehicle brake light detection using monocular vision | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/7247631/ | ||||||
2016 | Efficient Deep Models for Monocular Road Segmentation | Road Boundary Detection | 解像度・GPU種別処理時間比較あり。定量評価がかなりきっちりしている。 | 52ms, AP:88.85~92.86% | |||||||
2016 | LG, Korea University | Efficient Lane Detection Based on Spatiotemporal Images | Lane Detection,Road Boundary Detection | ピックアップ済み、自車線のみ、フレアやかすれに強い | 117ms, 79.6~96.8% | https://ieeexplore.ieee.org/document/7217838 | |||||
2016 | University of S˜ao Paulo | Exploiting Fully Convolutional Neural Networks for Fast Road Detection | Road Boundary Detection | CNN, 学習289枚 | KITTIでは2位、圧倒的に速いがVPGNetの50msと大差ない、2015Vision-Based...と全く同著者。歩道にFP。 | 32.7ms(Titan X), MaxF:90.79(KITTI40位レベル),Acc:93.9% | https://www.hds.utc.fr/~vfremont/dokuwiki/_media/en/icra16_1263_fi.pdf | ||||
2016 | Fast Traffic Sign Recognition Using Color Segmentation and Deep Convolutional Networks | TSR | CNN, SVM | , 89.71%~ | http://users.diag.uniroma1.it/bloisi/papers/bloisi-acivs2016-draft.pdf | ||||||
2016 | University of California San Diego | Looking at Vehicles in the Night: Detection & Dynamics of Rear Lights | Tail Light Detection,Turn Light Detection | VeDANt (Vehicle Detection using Active-learning during Nighttime) | tracking併用で精度向上 | 71ms(MATLAB), 92.3~98.5% | https://pdfs.semanticscholar.org/c8ac/5e08d8e2dfb66982de431f429b900248a8da.pdf | ||||
2016 | College of Engineering Sasthamcotta | Night Time Vehicle Detection Using Tail Lights: A Survey | Tail Light Detection | サーベイ、ボリューム少ない | , | https://pdfs.semanticscholar.org/3697/ea8c1c5bc35b1d4605a8e9ec629ade81eb86.pdf | |||||
2016 | Inha University | Night-time vehicle detection using low exposure video enhancement and lamp detection | Tail Light Detection | , | https://ieeexplore.ieee.org/document/7563005 | ||||||
2016 | University of Wollongong | Pedestrian lane detection in unstructured scenes for assistive navigation | Road Boundary Detection | , F-measure:95.3 | https://www.sciencedirect.com/science/article/pii/S1077314216000369 | ||||||
2016 | Chemnitz University of Technology | Road Detection through Supervised Classification | Road Boundary Detection | binary decision trees+v | 他手法との比較はなし | , maximum DR of 78% and 19.8% FPR | https://arxiv.org/pdf/1605.03150.pdf | ||||
2016 | Robust lane marking detection using boundary-based inverse perspective mapping | Lane Detection | 2値化→上下スキャン | IPM後に特徴抽出 | , | https://s3.amazonaws.com/academia.edu.documents/44111093/zqying_icassp2016_paper.pdf?AWSAccessKeyId=AKIAIWOWYYGZ2Y53UL3A&Expires=1558237095&Signature=NjAnTSjWNlcW%2Fm0bfInC59IpCXI%3D&response-content-disposition=inline%3B%20filename%3DROBUST_LANE_MARKING_DETECTION_USING_BOUN.pdf | |||||
2016 | Jilin University | Study on leading vehicle detection at night based on multisensor and image enhancement method | Tail Light Detection | OTSU,D-S evidence theory | fusion | , | http://downloads.hindawi.com/journals/mpe/2016/5810910.pdf | ||||
2016 | School of Science and Technology, Communication University of China | The research on traffic sign recognition based on deep learning | TSR | DBM+CCA | , | https://ieeexplore.ieee.org/document/7751612 | |||||
2016 | Software School of Xiamen University | Traffic flow detection based on the rear-lamp and virtual coil for nighttime conditions | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/7888317/ | ||||||
2016 | Traffic Sign Detection and Recognition using Features Combination and Random Forests | TSR | , 96.13% | https://thesai.org/Downloads/Volume7No1/Paper_93-Traffic_Sign_Detection_and_Recognition.pdf | |||||||
2016 | Traffic sign detection and recognition using fully convolutional network guided proposals | TSR | , 97.69 | http://www.vlrlab.net/admin/uploads/avatars/Traffic_sign_detection_and_recognition_using_fully_convolutional_network_guided_proposals.pdf | |||||||
2016 | Tsinghua University | Traffic-sign detection and classification in the wild | TSR | CNN | , | https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Zhu_Traffic-Sign_Detection_and_CVPR_2016_paper.pdf | |||||
2016 | VIT University | Vision Based vehicle detection: A literature review | Tail Light Detection | テールランプはほとんど記述なし | , | https://pdfs.semanticscholar.org/54d3/3eebacd199a9aca578f6402ddc85bf265a14.pdf | |||||
2016 | Aalborg Universitet | Vision for looking at traffic lights: Issues, survey, and perspectives | TLR | fuzzy clustering | サーベイ | , | http://vbn.aau.dk/ws/files/224024952/main.pdf | ||||
2016 | Nanjing University of Science and Technology | Vision-based two-step brake detection method for vehicle collision avoidance | Tail Light Detection | SVM | ブレーキ検知 | , | https://www.sciencedirect.com/science/article/pii/S0925231215012874 | ||||
2016 | Vel Tech University | Vision-based Vehicle Detection Survey | Tail Light Detection | , 記載なし | https://online-journals.org/onlinejour/index.php/i-jes/article/download/5590/3830 | ||||||
2015 | An Empirical Evaluation of Deep Learning on Highway Driving | Lane Detection | 44fps(GTX 780Ti), | https://arxiv.org/pdf/1504.01716.pdf | |||||||
2015 | A Reliable Method for Detecting Road Regions from a Single Image Based on Color Distribution and Vanishing Point Location | Road Boundary Detection | 消失点推定、路面のアプリオリ情報が必要 | , | https://pdf.sciencedirectassets.com/280203/1-s2.0-S1877050915X00202/1-s2.0-S1877050915021134/main.pdf?x-amz-security-token=AgoJb3JpZ2luX2VjEDQaCXVzLWVhc3QtMSJGMEQCIE%2BifPDAGZwVUtkAbXhtS2MzhEBiueXc%2FsOq5eJb1MPMAiBELH6lTT0AkQ4G%2Flb1W8j2akp%2BOQnXKX%2BToXN4ylH3lyraAwg8EAIaDDA1OTAwMzU0Njg2NSIMvWqdcTeKy6R0LrWiKrcD%2Bc5MQb%2FKDrok4FxfAXi2YJDP6iOrxQvHC35GZKEzs1KkMTg0EIc3ssoBvU4QUIPYVdGVAS9OlMQM%2BleBiUEfEaNljg1zDTncI%2F63hG5d3fyT%2FhYag8fQd1c0HLb33uYp%2B4oT5uKmbTHk%2FAsC5uLxY0Uc3kvdToMjVQwVEPhn6OzjBZTczytVcwopLix%2FXCeQt5PDpL8RW3Iteu0Z9mavStxUF6WCJyomUc06C%2F4uQwwmlci81zslg7ZAcqq8BzMqlMwxf16u2uLcSgVoaW96LHhdsjYj8A74Vpd%2F%2B%2Fv7y4ZbSdvfb9U1GU3ZAj%2BllsJEiPi6gdiiMylsWcCPEVHX7PlwSp%2BQj3QG%2FXdJolwkn1crg2CqJArZBodrCwNRTW9DlIpIdDwAGTPRPFk1laxwuhlp6ziLOot2yyVUOo8E%2FwbGDia4sxYHAO1p5t1t1GAlDLElJuFeaZQuvhoGhwyTt2iLDL1vKkbut6%2Fo1QLVfDL1r%2Bj5hD230TVaFhyw0oLeK7EVQRNmPbeKn1nlYYWB54zWUIsek4bQE3JEyJDPK2J1BEC%2F2iEFE3NUDTNRxeJ9N9keFV4ZrDC48v3mBTq1Aai2aiBvaG8c62E2LE399xYSc%2BWPbQDyexylwJ%2F%2FtZDkRcOPjBNQrPbBLN8YGQGU4anUgToVCmr4Wm4QKpxKRLklLcyrWKY0GrRGUrP9SdSfSzKGr0j6T99SkwnsDEvLO%2BkhW6m45zrnD9nwKcF9yrXn8ggdrHUjA%2BruiFetBuhER4lv7xJ1woHnKl1oAPAoNGCCnqRh9POsn5itu5h3rvKnK8L3kKYR6pqVo%2FO4J0ZEjb8d7nQ%3D&AWSAccessKeyId=ASIAQ3PHCVTY7PZPJNF5&Expires=1558154596&Signature=vb%2BTBHTOsNWYDwYW%2B9cEVs0nQKs%3D&hash=c6366454fc7fb2d8a5412105f3a69f69453f22f6811fed7789e10db95e863dc3&host=68042c943591013ac2b2430a89b270f6af2c76d8dfd086a07176afe7c76c2c61&pii=S1877050915021134&tid=spdf-d1af04de-2c24-4cca-9578-d7595d4a0901&sid=d83c7e88325cc043bd1a8aa70578201085a9gxrqb&type=client | ||||||
2015 | IIT Guwahati | A Robust Lane Detection and Departure Warning System | Lane Detection | RANSAC, optical flow, GMM | 速度で[21]の方が有利に見える→論文見るとカーブが全然だめ(BL) | 29ms(SPRAYより速いが精度は劣る、遠距離ほど他手法より有利), 94.25~98.44% | https://arxiv.org/pdf/1504.07590.pdf | ||||
2015 | Universidad de las Palmas de Gran Canaria | A Survey on Traffic Light Detection | TLR | , | https://www.researchgate.net/publication/281286347_A_Survey_on_Traffic_Light_Detection | ||||||
2015 | A Survey on Traffic Light Detection | TLR | , | https://www.researchgate.net/publication/281286347_A_Survey_on_Traffic_Light_Detection | |||||||
2015 | Peking University | A vision-based hierarchical framework for autonomous front-vehicle taillights detection and signal recognition | Turn Light Detection | 昼間対応 | , | https://ieeexplore.ieee.org/abstract/document/7313248 | |||||
2015 | Computer Vision at Intersections Explorations in Driver Assistance Systems and Data Reduction for Naturalistic Driving Studies | TLR | 2016Vision for looking at...を含む修論 | , | https://projekter.aau.dk/projekter/files/213565236/main.pdf | ||||||
2015 | Computer Vision at Intersections Explorations in Driver Assistance Systems and Data Reduction for Naturalistic Driving Studies | TLR | ACF | Piotr’s Matlab Toolboxの利用の詳細が記載 | , | https://projekter.aau.dk/projekter/files/213565236/main.pdf | |||||
2015 | Fast pixelwise road inference based on Uniformly Reweighted Belief Propagation | Road Boundary Detection | PGM-ARS | KITTIでは62位、SPRAYと同等、影のFN、車のFPが問題 | <50ms(i7-4700MQ), MaxF:85.52%(SPRAYは86.33%) | http://www.robesafe.com/personal/bergasa/papers/MarioPassani_IV2015.pdf | |||||
2015 | Fast Symbolic Road Marking and Stop-line Detection for Vehicle Localization | Road Marking Recognition | total error rate (TER)-based classifier, HOG | ダントツで高速、多クラス、 | 4.5ms, 99.2% | http://web.yonsei.ac.kr/hgjung/Ho%20Gi%20Jung%20Homepage/Publications/2015/IVS2015(Fast%20Symbolic%20Road%20Marking%20and%20Stop-line%20Detection%20for%20Vehicle%20Localization).pdf | |||||
2015 | FPGA Based Traffic Sign Detection for Automotive Camera Systems | TSR | 下記サーベイで引用 | 60fps@full HD, | http://radio-project.eu/downloads/publications/schwiegelshohn_etal-2015.pdf | ||||||
2015 | Shandong University | Nighttime Vehicle Detection for Heavy Trucks | Tail Light Detection | AdaBoost,Improved Threshold | , 89.4% | https://www.researchgate.net/profile/Hui_Chen56/publication/288128551_Nighttime_Vehicle_Detection_for_Heavy_Trucks/links/567e606608aebccc4e055156/Nighttime-Vehicle-Detection-for-Heavy-Trucks.pdf | |||||
2015 | Parul Institute of Engineering and Technology | Nighttime vehicle Tail light detection in low light video frames using Matlab | Tail Light Detection | 60.64~88.29%, | https://www.ijraset.com/fileserve.php?FID=2307 | ||||||
2015 | Aalborg University | Ongoing Work on Traffic Lights | TLR | , | http://vbn.aau.dk/files/217928577/mainAVSS.pdf | ||||||
2015 | University of Oxford | Reading the Road: Road Marking Classification and Interpretation | Road Marking Recognition | CRF, Delaunay triangulation | , 73.6~93.2% | http://www.robots.ox.ac.uk/~mobile/Papers/road-marking-classification-paul-ingmar.pdf | |||||
2015 | Real-time illumination invariant lane detection for lane departure warning system | Lane Detection | 適応ROI | ラインの色の考察 | , | http://diml.yonsei.ac.kr/papers/Real-time%20Illumination%20Invariant%20Lane%20Detection%20%20for%20Lane%20Departure%20Warning%20System.pdf | |||||
2015 | Universitat Munchen | Real-time lane detection and tracking on high performance computing devices | Road Boundary Detection | 画像を見ると精度はいまいち | , | http://www.i6.in.tum.de/pub/Main/Hub/detection.pdf | |||||
2015 | University of Kerala Kariavattom | Review of lane detection and tracking algorithms in advanced driver assistance system | Road Boundary Detection | サーベイ | , | https://s3.amazonaws.com/academia.edu.documents/38630775/7415ijcsit06.pdf?AWSAccessKeyId=AKIAIWOWYYGZ2Y53UL3A&Expires=1553137239&Signature=BlYCfvTmsHKj%2FugRUNXKAOKEFiQ%3D&response-content-disposition=inline%3B%20filename%3DREVIEW_OF_LANE_DETECTION_AND_TRACKING_AL.pdf | |||||
2015 | Prescott College of Engineering | Robust and computationally lightweight autonomous tracking of vehicle taillights and signal detection by embedded smart cameras | Head Light Detection,Turn Light Detection | ヘッドライト拡張可 | No.70と同著者同内容 | 123ms/台,>300ms @ 624MHzMPU, | https://ieeexplore.ieee.org/abstract/document/7031905/ | ||||
2015 | Robust and computationally lightweight autonomous tracking of vehicle taillights and signal detection by embedded smart cameras | Tail Light Detection,Turn Light Detection | ルールベース,LKF | 同著者、ヘッドライト、昼夜可 | , B:74.4~98.6%,T:68.4~100%,H:91.9% | https://ieeexplore.ieee.org/abstract/document/7031905/ | |||||
2015 | Beijing Jiaotong University | Robust nighttime vehicle detection by tracking and grouping headlights | Head Light Detection | , Tracking:86.4% | https://slideplayer.com/slide/9261334/ | ||||||
2015 | Universidad de las Palmas de Gran Canaria, | Robust real-time traffic light detection and distance estimation using a single camera | TLR | 2016Vision for looking at traffic lights: Issues, survey, and perspectivesでトップ | , 99.4% | https://www.researchgate.net/publication/272240041_Robust_real-time_traffic_light_detection_and_distance_estimation_using_a_single_camera | |||||
2015 | Lee-Ming Institute of Tech. | Robust rear light status recognition using symmetrical surfs | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/7313424/ | ||||||
2015 | Islamic Azad University | Robust Vehicle Detection and Distance Estimation Under Challenging Lighting Conditions | Tail Light Detection | 単眼測距、上記引用、混雑条件、距離依存への対策、昼夜兼用 | 25-28fps(CPU), Prec:95.1% | https://ieeexplore.ieee.org/document/7101268 | |||||
2015 | Aalborg University | Traffic Light Detection: A Learning Algorithm and Evaluations on Challenging Dataset | TLR | , | http://vbn.aau.dk/ws/files/234499286/mainITSC.pdf | ||||||
2015 | Universiti Teknologi Petronas | Vehicle Detection Techniques for Collision Avoidance Systems: A Review | Tail Light Detection | サーベイ、2011A Real-Time Visionを引用 | , | https://www.researchgate.net/profile/Amir_Mukhtar2/publication/277574813_Vehicle_Detection_Techniques_for_Collision_Avoidance_Systems_A_Review/links/56776f5d08ae125516ec0dbc/Vehicle-Detection-Techniques-for-Collision-Avoidance-Systems-A-Review.pdf | |||||
2015 | Information College of Beijing Union University | Vision-Based Method for Forward Vehicle Brake Lights Recognition | Tail Light Detection | 雨天、昼間の過酷条件で評価 | , 80.01~85.62%,赤い車だと50.31~55.25% | https://www.researchgate.net/publication/286523979_Vision-Based_Method_for_Forward_Vehicle_Brake_Lights_Recognition | |||||
2015 | University of S˜ao Paulo | Vision-Based Road Detection using Contextual Blocks | Road Boundary Detection | KITTIのnon-DL/LiDARでHIMに次ぐ(46位) | 2s, AP:75.21%~ | https://arxiv.org/pdf/1509.01122.pdf | |||||
2014 | University Hadj Lakhdar | A Fast and Robust Traffic Sign Recognition | TSR | 検出限定で最速、他の対象に応用できないか? | 検出のみ<1ms(360x270), 95.65%(detection) | https://pdfs.semanticscholar.org/f689/534a9d6af365e7628a66bd91c7658d02b932.pdf | |||||
2014 | University of Zagreb | A Method for On-road Night-time Vehicle Headlight Detection and Tracking | Head Light Detection,Tail Light Detection | JPDAF | 距離・速度推定 | , Tracking:87/92% | https://bib.irb.hr/datoteka/741339.A_Method_for_On-road_Night-time_Vehicle_Headlight_Detection_and_Tracking.pdf | ||||
2014 | University of Bedfordshire | An algorithm for accurate taillight detection at night | Tail Light Detection | Rule-based(color filtering→pairing→symmetry check) | 854ms, 95.37% | http://uobrep.openrepository.com/uobrep/bitstream/10547/333824/2/taillight.pdf | |||||
2014 | University "Parthenope" Naples | An automated nighttime vehicle counting and detection system for traffic surveillance | Head Light Detection | adaptive threshold, contour detection | バイクも検出可能 | real time(320x180), Precision:98.71% | https://www.researchgate.net/profile/Giuseppe_Salvi/publication/265172351_An_Automated_Nighttime_Vehicle_Counting_and_Detection_System_for_Traffic_Surveillance/links/55af466308aed614b09a8627/An-Automated-Nighttime-Vehicle-Counting-and-Detection-System-for-Traffic-Surveillance.pdf | ||||
2014 | Norwich University | Autonomous tracking of vehicle taillights and alert signal detection by embedded smart cameras | Turn Light Detection | 2015Robust and computationally ...と同著者、昼夜兼用がポイント、疑似コードあり | 115~758ms(MATLAB), 90.3~98.6%(昼、多車線の方が不利) | https://www.researchgate.net/profile/Emmanuel_Tonye/publication/272161147_A_Parallel_Approach_for_Statistical_Texture_Parameter_Calculation/links/54dc6d2c0cf2a7769d961a5a.pdf#page=131 | |||||
2014 | DENSO | Car detection at night using latent filters | Tail Light Detection | 117ms(VGA), | https://www.researchgate.net/profile/Hossein_Tehrani3/publication/269293903_Car_detection_at_night_using_latent_filters/links/5704703b08ae74a08e24644c.pdf | ||||||
2014 | Universit´e de Technologie de Compi`egne | Comprehensive performance analysis of road detection algorithms using the common urban Kitti-road benchmark | Road Boundary Detection | HistonBoost(multi-normalized-histogram joint boosting[21]),ANN | ベンチマーキング、定量評価手法 | , | https://hal.archives-ouvertes.fr/hal-01089094/document | ||||
2014 | Detection and Recognition of Road Traffic Signs in Various Illumination and Weather Conditions | TSR | , 89.47 | https://www.ijert.org/research/detection-and-recognition-of-road-traffic-signs-in-various-illumination-and-weather-conditions-IJERTV3IS071337.pdf | |||||||
2014 | Bielefeld University | Monocular Road Terrain Detection by Combining Visual and Spatial Information | Road Boundary Detection | SPRAY | ノイジーでFPが多い | ~45ms(GTX 580), AP:95.6(road area), 85.2(ego-lane) | https://www.honda-ri.de/pubs/pdf/1931.pdf | ||||
2014 | Jilin University | Multifeature Fusion Vehicle Detection Algorithm Based on Choquet Integral | Tail Light Detection | 車両検出なので除外 | 50ms(896 × 592), | https://www.hindawi.com/journals/jam/2014/701058/ | |||||
2014 | Amrita School of Engineering | Night Time Vehicle Detection for Real Time Traffic Monitoring Systems A Review | Tail Light Detection | Rule Based/Symmetry Based/SVM/Hypothesis Generationを比較 | , Rule Basedが98%でベスト | https://pdfs.semanticscholar.org/8724/baf615c79a1ee5b9dbd1c2ee9d1b3a16b46f.pdf | |||||
2014 | Yuan Ze University | Nighttime turn signal detection by scatter modeling and reflectance-based direction recognition | Turn Light Detection | Nakagami imagingで不変量抽出, AdaBoost | 遠い方が検出率が高い | , 76~87% | http://mmplab.cs.ntou.edu.tw/Publication/Paper_File/A_National_Journal/2014/Nighttime%20Turn%20Signal%20Detection%20by%20Scatter%20Modeling%20and%20Reflectance-based%20Direction%20Recognition.pdf | ||||
2014 | University of California San Diego | On Performance Evaluation Metrics for Lane Estimation | Road Boundary Detection | ELAS論文の引用 | , | http://cvrr.ucsd.edu/publications/2014/SatzodaTrivedi_ICPR2014.pdf | |||||
2014 | University of Zagreb | On-road Night-time Vehicle Light Detection and Tracking Methods Overview | Head Light Detection,Tail Light Detection | , 距離との関係まで言及(Headlights are detected at distances between 300 and 500 m and tail-lights at distances between 30 - 50 m...) 300m以内のHL,30-50mのTLで100% | https://pdfs.semanticscholar.org/d5cb/d2a163cabcb2e3b62b440ea639f01ab50eb7.pdf | ||||||
2014 | Xi'an Institute of High-Tech | Preceding vehicle detection and tracking adaptive to illumination variation in night traffic scenes based on relevance analysis | Tail Light Detection | 14ms(768 × 576), 67~98.8% | https://www.mdpi.com/1424-8220/14/8/15325/pdf | ||||||
2014 | California Institute of Technology | Real time Detection of Lane Markers in Urban Streets | Road Marking Recognition | 車線のみ。データセット、スクリプト公開。 | , | https://arxiv.org/pdf/1411.7113.pdf | http://www.mohamedaly.info/software/caltech-lane-detection | ||||
2014 | Worcester Polytechnic Institute | Real-Time Traffic Sign Detection and Recognition Using GPU | TSR | HOG+SVM | window size of 32 by 32 Tesla K20で13~17 ms | , | http://www.ieee-hpec.org/2014/CD/index_htm_files/FinalPapers/109.pdf | ||||
2014 | Institute of Automation, Chinese Academy of Sciences | Rear-View Vehicle Detection and Tracking by Combining Multiple Parts for Complex Urban Surveillance | Tail Light Detection | 最速、License plateも同時、定量評価が充実、天候によりる精度の変化も明確 | 25ms(1920*1072以上), Recall:90.3~99.1%(daytimeのprecisionは73.5%まで下がる) | http://ir.ia.ac.cn/bitstream/173211/3639/1/Rear-View%20Vehicle%20Detection%20and%20Tracking%20by%20Combining%20Multiple%20Parts%20for%20Complex%20Urban%20Surveillance.pdf | |||||
2014 | General Motors | Recent Progress in Road and Lane Detection - A survey | Road Boundary Detection | , | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.233.1475&rep=rep1&type=pdf | ||||||
2014 | Robust Lane Detection in Shadows and Low Illumination Conditions using Local Gradient Features | Lane Detection | , 95.54% | https://pdfs.semanticscholar.org/1b0d/6986045c69b69c86cf2fbefe5ea60da29aff.pdf | |||||||
2014 | Yuan Ze University | Robust nighttime turn signal direction recognition | Turn Light Detection | 2014Nighttime turn signal...と同内容ぽい | , >80% | https://ieeexplore.ieee.org/iel7/6894724/6903991/06904059.pdf | |||||
2014 | Robust real-time traffic light detection and distance estimation using a single camera | TLR | 疑似コードあり、参考になるノウハウも多そう | , | https://www.researchgate.net/publication/272240041_Robust_real-time_traffic_light_detection_and_distance_estimation_using_a_single_camera | ||||||
2014 | Mokpo National University | State Machine and Downhill Simplex Approach for Vision-Based Nighttime Vehicle Detection | Head Light Detection | , 94.2% | https://onlinelibrary.wiley.com/doi/pdf/10.4218/etrij.14.0113.0509 | ||||||
2014 | National Central University(Taiwan) | Vehicle color classification using manifold learning methods from urban surveillance videos | NFL (20) + SVM (RBF-kernel function) | 昼間 | 18ms, 88.18 (±0.89) | https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/1687-5281-2014-48 | |||||
2014 | University of California San Diego | Vision-based Lane Analysis Exploration of Issues and Approaches for Embedded Realization | Lane Detection | LASeR | 上記手法の計算量解析 | 記載なし, | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.649.1011&rep=rep1&type=pdf | ||||
2014 | Fraunhofer Institute | Vision-Based Robust Road Lane Detection in Urban Environments | Lane Detection,Road Boundary Detection | 影が除去できる。車線検出は曲線対応で遠方まで可能。逆光の照り返しで失敗。 | , F:93.33% | https://www.researchgate.net/profile/Michael_Beyeler2/publication/269519254_Vision-based_robust_road_lane_detection_in_urban_environments/links/552bfc2e0cf21acb091f4263.pdf | |||||
2014 | Daimler | Will this car change the lane?-turn signal recognition in the frequency domain | Turn Light Detection | , | https://ieeexplore.ieee.org/abstract/document/6856477/ | ||||||
2013 | Honda Research Institute Europe | A new performance measure and evaluation benchmark for road detection algorithms | Road Boundary Detection | SPRAY,CNN | SPRAY論文の著者による定量評価方法提案(pixel-based/behavior-based)。2015A Robust Lane Detection...でベンチマーク。BaseLineは近距離だけなら速度で圧倒的有利だが、アルゴリズムの詳細がない。 | 20ms(1core,Python), | http://www.cvlibs.net/publications/Fritsch2013ITSC.pdf | ||||
2013 | Tsinghua University | A Region Tracking-Based Vehicle Detection Algorithm in Nighttime Traffic Scenes | Tail Light Detection | 142ms(MATLAB), 97.472%,93.45~100% | https://www.mdpi.com/1424-8220/13/12/16474/pdf | ||||||
2013 | A Sensor-Fusion Drivable-Region and Lane-Detection System for Autonomous Vehicle Navigation in Challenging Road Scenarios | Lane Detection | LiDAR | , >91.2% | https://www.researchgate.net/profile/Andreas_Nuchter/publication/260522418_A_Sensor-Fusion_Drivable-Region_and_Lane-Detection_System_for_Autonomous_Vehicle_Navigation_in_Challenging_Road_Scenarios/links/5890ba0b92851cda25689d4c/A-Sensor-Fusion-Drivable-Region-and-Lane-Detection-System-for-Autonomous-Vehicle-Navigation-in-Challenging-Road-Scenarios.pdf | ||||||
2013 | University of Bochum | Detection of Traffic Signs in Real-World Images The German Traffic Sign Detection Benchmark | TSR | detectionではViola-Jonesがベスト | , ~100% | https://www.researchgate.net/publication/242346625_Detection_of_Traffic_Signs_in_Real-World_Images_The_German_Traffic_Sign_Detection_Benchmark | |||||
2013 | Hanyang University | Enhancing Light Blob Detection for Intelligent Headlight Control Using Lane Detection | Head Light Detection | 低露出+自動露出(車線検出・tailの遠距離用), fixed threshold | ROC curves, 画像の定性評価あり | , ROC曲線のみ | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.310.867&rep=rep1&type=pdf | ||||
2013 | Uludag University | Keeping the Vehicle on the Road – A Survey on On-Road Lane Detection Systems | Road Boundary Detection | 2000年代の論文の比較 | , | http://romisatriawahono.net/lecture/rm/survey/computer%20vision/Yenikaya%20-%20Keeping%20the%20Vehicle%20on%20the%20Road%20-%202013.pdf | |||||
2013 | Hunan University | Lane Detection Algorithm based on Local Feature Extraction | Lane Detection | Median Local Threshold | Rainyの方が精度が高い | <200ms(MATLAB,256 × 240), 87.36~98.67% | https://www.researchgate.net/profile/Shutao_Li/publication/271469116_Lane_detection_algorithm_based_on_local_feature_extraction/links/5639a7c808aecf1d92a9c817.pdf | ||||
2013 | University of California, San Diego | Looking at Vehicles on the Road: A Survey of Vision-Based Vehicle Detection, Tracking, and Behavior Analysis | Tail Light Detection | 車両検出・追尾・挙動解析のサーベイ | , | http://swiftlet.ucsd.edu/publications/2013/SayananTrivedi_IEEETITS2013_VehicleSurvey.pdf | |||||
2013 | Night Time Car Recognition Using MATLAB | Tail Light Detection | , | http://www.cipa.dcu.ie/papers/eProject_Hazim%20Hamza_2013.pdf | |||||||
2013 | Night-Time Traffic Light Detection Based On SVM with Geometric Moment Features | TLR | SVM | , | https://waset.org/publications/8944/night-time-traffic-light-detection-based-on-svm-with-geometric-moment-features | ||||||
2013 | Xi’an Technology University | Nighttime vehicle detection algorithm based on brightness cumulative histogram | Tail Light Detection | , | http://en.cnki.com.cn/Article_en/CJFDTotal-JSJC201306055.htm | ||||||
2013 | Real-time detection and classification of road lane | Lane Detection,Road Boundary Detection | ラインの分類可能。 | 25ms @VGA, 71.53~99.32% | https://www.researchgate.net/publication/261022358_Real-Time_Detection_and_Classification_of_Road_Lane_Markings/download | ||||||
2013 | University of California, San Diego | Selective Salient Feature based Lane Analysis | Lane Detection | LASeR | On Performance Evaluation...で引用 | 記載なし, 90-95%, <8cm(自車位置) | http://swiftlet.ucsd.edu/publications/2013/SatzodaLASER13.pdf | ||||
2013 | University of California, San Diego | Towards Automated Drive Analysis: A Multimodal Synergistic Approach | Lane Detection | Selective Salient...の評価方法の引用 | , | http://cvrr.ucsd.edu/publications/2013/SatzodaDriveAnalysis13.pdf | |||||
2013 | Universitas Indonesia | Vehicle counting and speed measurement using headlight detection | Head Light Detection | 定点監視、速度推定 | , | https://www.researchgate.net/profile/Adi_Nurhadiyatna/publication/271548723_Vehicle_counting_and_speed_measurement_using_headlight_detection/links/55273ab40cf229e6d63613e1.pdf | |||||
2013 | Chinese Academy of Sciences | Vehicle detection based on the and–or graph for congested traffic conditions | Tail Light Detection | and -or graph (AOG) | occlusion,時刻、天候の変動に強い | , | https://ieeexplore.ieee.org/abstract/document/6480875/ | ||||
2013 | Versatile lane departure warning using 3D visual geometry | Lane Detection | lateral inhibition property | , | http://www.ijicic.org/ijicic-12-02077.pdf | ||||||
2012 | Jawaharlal Nehru Technological University(India) | A computer vision model for vehicle detection in traffic surveillance | Head Light Detection | , 87.666~95.937% | http://www.ijesat.org/Volumes/2012_Vol_02_Iss_05/IJESAT_2012_02_05_05.pdf | ||||||
2012 | Tsinghua University | A Global and Local Condensation for Lane Tracking | Lane Detection | particle filter(particle数5000) | ライントラッキング、ピッチ変化考慮でIPM精度向上 | , | https://www.researchgate.net/profile/Zhidong_Deng/publication/261202026_A_global_and_local_condensation_for_lane_tracking/links/58e45c0d4585159f7a777d5b/A-global-and-local-condensation-for-lane-tracking.pdf | ||||
2012 | Tianjin Polytechnic University | A Night Time Application for a Real-Time Vehicle Detection Algorithm Based on Computer Vision | Head Light Detection | Rule-based | 64fps(720×576), 97.93% | https://pdfs.semanticscholar.org/0c1e/f97a45ae19066755b20d9c39f3c10337af28.pdf | |||||
2012 | Syracuse University | A robust algorithm for the detection of vehicle turn signals and brake lights | Tail Light Detection,Turn Light Detection | , | https://ieeexplore.ieee.org/abstract/document/6328045/ | ||||||
2012 | A robust algorithm for the detection of vehicle turn signals and brake lights | Tail Light Detection,Turn Light Detection | , | https://ieeexplore.ieee.org/abstract/document/6328045/ | |||||||
2012 | Syracuse University | Autonomous tracking of vehicle rear lights and detection of brakes and turn signals | Tail Light Detection,Turn Light Detection | 上と同著者 | , | https://ieeexplore.ieee.org/document/6291543 | |||||
2012 | Clemson University | Brake lamp detection in complex and dynamic environments: Recognizing limitations of visual attention and perception | Tail Light Detection | シミュレータで実験 | , | http://andrewd.ces.clemson.edu/research/vislab/docs/AAP45(2012).pdf | |||||
2012 | Yuan Ze University | Frequency-tuned taillight-based nighttime vehicle braking warning system | Tail Light Detection | 上記引用、Nakagami imagingを使った論文を複数書いている著者 | , | https://ieeexplore.ieee.org/document/6265338 | |||||
2012 | Ruhr-Universität Bochum | Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition | TSR | , | |||||||
2012 | TEI, RAJWADA | Night Time Vehicle Detection and Classification Using Support Vector Machine | Head Light Detection,Tail Light Detection | SVM | 15~30fps, | https://pdfs.semanticscholar.org/ac25/b7c5e7cd65f785c2e2e20ed881850253125f.pdf | |||||
2012 | Hunan University | Night-time vehicle detection using DS evidence theory | Tail Light Detection | TEI, RAJWADA | , | http://en.cnki.com.cn/Article_en/CJFDTOTAL-JSYJ201205092.htm | |||||
2012 | Yuan Ze University | Nighttime brake-light detection by Nakagami imaging | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/6210387/ | ||||||
2012 | Chaoyang University of Technology | PSO algorithm particle filters for improving the performance of lane detection and tracking systems in difficult roads | Lane Detection | , 記載なし | https://www.mdpi.com/1424-8220/12/12/17168/pdf | ||||||
2012 | Real-Time Detection and Recognition of Road Traffic Signs | TSR | , | http://www.cs.sjtu.edu.cn/~shengbin/course/SE/Real-Time%20Detection%20and%20Recognition%20of%20road%20traffic%20signs.pdf | |||||||
2012 | Chinese Academy of Sciences | Rear lamp based vehicle detection and tracking for complex traffic conditions | Tail Light Detection | MSER | ノイズにロバスト | , | https://ieeexplore.ieee.org/abstract/document/6204950/ | ||||
2012 | University of Patras | Rear lights vehicle detection for collision avoidance | Tail Light Detection | Otsuの閾値処理→pairing→エッジ検出→対称性チェック | 車幅から距離推定 | , 92.6% | https://www.researchgate.net/profile/George_Siogkas/publication/230601632_Rear_Lights_Vehicle_Detection_for_Collision_Avoidance/links/004635318f8c882741000000/Rear-Lights-Vehicle-Detection-for-Collision-Avoidance.pdf | ||||
2012 | General Motors | Recent progress in road and lane detection: a survey | Road Boundary Detection | サーベイ、文章のみで辛い | , | https://sites.google.com/site/danmlevi/LaneRoadSurvey_final.pdf | |||||
2012 | Yuan-Ze University | Salient video cube guided nighttime vehicle braking event detection | Tail Light Detection | , | https://www.sciencedirect.com/science/article/pii/S1047320312000260 | ||||||
2012 | Bielefeld University | Spatial Ray Features for Real-Time Ego-Lane Extraction | Road Boundary Detection | , | http://www.cvlibs.net/projects/autonomous_vision_survey/literature/Kuehnl2012IV.pdf | ||||||
2012 | Huazhong University of Science and Technology | Tracking and Pairing Vehicle Headlight in Night Scenes | Head Light Detection | ROC評価あり。bounding boxは出さない。 | , Tracking:88.2/95.2% | http://www.cvsslab.com/publication/67.pdf | |||||
2012 | Traffic sign detection using computer vision Explorations for a driver support system | TSR | , | https://www.academia.edu/9716924/Traffic_sign_detection_using_computer_vision_Explorations_for_a_driver_support_system?auto=download | |||||||
2012 | North China University of Technology | Vehicle Detection and Tracking at Night in Video Surveillance | Tail Light Detection | 車両BBを出力 | , 97.07% | https://online-journals.org/index.php/i-joe/article/viewFile/2828/2683 | |||||
2012 | Chengdu University of Information and Technology | Vehicle detection based on two-way multilane at night | Tail Light Detection | , | http://en.cnki.com.cn/Article_en/CJFDTOTAL-TXJS201210020.htm | ||||||
2012 | Future University Hakodate | Vehicles detection based on extremas in nighttime driving scene | Head Light Detection,Tail Light Detection | ヘッドライト兼用(入力がgray/red画像かの違い) | , | https://ieeexplore.ieee.org/abstract/document/6379950/ | |||||
2012 | Aalborg University | Vision-Based Traffic Sign Detection and Analysis for Intelligent Driver Assistance Systems: Perspectives and Survey | TSR | サーベイ | , | http://vbn.aau.dk/files/72574480/survey.pdf | |||||
2012 | Yuan Ze University | Visual-Based Spatiotemporal Analysis for Nighttime Vehicle Braking Event Detection | Tail Light Detection | 時系列からブレーキ検知 | , | https://link.springer.com/chapter/10.1007/978-3-642-27355-1_81 | |||||
2011 | National Taipei University of Technology | A Real-Time Vision System for Nighttime Vehicle Detection and Traffic Surveillance | Head Light Detection,Tail Light Detection | segmentation | tail lampも対象、バイクも検出可能,比較している他手法との違いが歴然 | 12ms, 96.3~97.9% | https://ir.nctu.edu.tw/bitstream/11536/8964/1/000289478000054.pdf | ||||
2011 | National Central University(Taiwan) | A vision-based system for the prevention of car collisions at night | Tail Light Detection | 3層3neuronFCのNN | 幾何的heuristicの詳細な説明、距離推定 | 32ms(720x480), 74.2~98.3% | https://www.researchgate.net/profile/Ming-Chih_Lu/publication/220465050_A_vision-based_system_for_the_prevention_of_car_collisions_at_night/links/0c96051ae9e1891e9c000000.pdf | ||||
2011 | Industrial Technology Research Institute(Taiwan) | Adaptive IPM-based lane filtering for night forward vehicle detection | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/5975840/ | ||||||
2011 | University of Alcal´a | Automatic LightBeam Controller for driver assistance | Tail Light Detection | 2014On-road Night-time...の[4], BB出力, 距離推定, High/Low beam推定 | ~20fps, 1 | https://www.researchgate.net/publication/225715555_Automatic_LightBeam_Controller_for_driver_assistance | |||||
2011 | North Carolina State University | Detection of multiple preceding cars in busy traffic using taillights | Tail Light Detection | 昼夜兼用 | , | https://link.springer.com/chapter/10.1007/978-3-642-21596-4_34 | |||||
2011 | Effective Traffic Lights Recognition Method for Real Time Driving Assistance System in the Daytime | TLR | , | https://waset.org/publications/725/effective-traffic-lights-recognition-method-for-real-time-driving-assistance-systemin-the-daytime | |||||||
2011 | 名大村瀬研 | Intelligent Traffic Sign Detector Adaptive Learning Based on Online Gathering of Training Samples | TSR | Real AdaBoost、Mean shift clustering、オンライン学習 | , | http://bird1.murase.m.is.nagoya-u.ac.jp/~ide/res/paper/E11-conference-ddeguchi-1pub.pdf | |||||
2011 | Georgia Institute of Technology | Multi-viewpoint lane detection with applications in driver safety systems | Lane Detection | ラインの定量評価方法 | , | https://smartech.gatech.edu/bitstream/handle/1853/43752/borkar_amol_a_201205_phd.pdf | |||||
2011 | Universitat Autònoma de Barcelona | Multiple-Target Tracking for Intelligent Headlights Control | Tail Light Detection | BELIEF PROPAGATION | pptあり | , Split and Mergeが52.8% | http://www.cvc.uab.es/~joans/journals/12%20ITS%20Multiple%20target%20tracking%20for%20intelligent%20headlights%20control.pdf | ||||
2011 | EPFL | Real-time vehicle tracking for driving assistance | Tail Light Detection | HSVでのHeadlight,Taillight,Blinkerの分離 | 39~46fps(720 × 576), 94.4~98.2% | http://doc.rero.ch/record/321782/files/138_2009_Article_243.pdf | |||||
2011 | Universitat Autònoma de Barcelona | Road Detection Based on Illuminant Invariance | Lane Detection,Road Boundary Detection | 影対策に有用、上記引用 | , | https://www.researchgate.net/publication/220108942_Road_Detection_Based_on_Illuminant_Invariance | |||||
2011 | Road sign detection and recognition system for real-time embedded applications | TSR | 昼夜比較 | , 夜は-40% | https://www.researchgate.net/publication/224247496_Road_sign_detection_and_recognition_system_for_real-time_embedded_applications | ||||||
2011 | Shanghai Jiao Tong University | Robust and real-time traffic lights recognition in complex urban environments | TLR | 40ms, 96.7% | https://download.atlantis-press.com/article/2436.pdf | ||||||
2011 | Ruhr-Universität Bochum | The German Traffic Sign Recognition Benchmark: A multi-class classification competition. | TSR | , | |||||||
2011 | Linköping University | Using Fourier Descriptors and Spatial Models for Traffic Sign Recognition | TSR | , | https://www.cvl.isy.liu.se/research/datasets/traffic-signs-dataset/SCIA2011_poster.pdf | ||||||
2011 | University of Missouri-Columbia | Vehicle detection algorithm based on light pairing and tracking at nighttime | Head Light Detection | , 88.2~98.8% | https://www.researchgate.net/profile/Weibing_Wan/publication/258496559_Vehicle_detection_algorithm_based_on_light_pairing_and_tracking_at_nighttime/links/573ae62c08ae9ace840e68cc/Vehicle-detection-algorithm-based-on-light-pairing-and-tracking-at-nighttime.pdf | ||||||
2011 | National University of Ireland | Vision-based detection and tracking of vehicles to the rear with perspective correction in low-light conditions | Tail Light Detection | , | https://digital-library.theiet.org/content/journals/10.1049/iet-its.2010.0032 | ||||||
2011 | Toyota Technological Institut | Vision-Based Vehicle Detection for Nighttime with Discriminately Trained Mixture of Weighted Deformable Part Models | Tail Light Detection | , 78.93% | https://www.researchgate.net/profile/Hossein_Tehrani3/publication/261151134_Vision-based_vehicle_detection_for_nighttime_with_discriminately_trained_mixture_of_weighted_deformable_part_models/links/57049ece08ae13eb88b68fbe.pdf | ||||||
2010 | An Adaptive Method for Lane Marking Detection Based on HSI Color Model | Lane Detection | , | https://www.researchgate.net/profile/Trung-Thien_Tran/publication/220777443_An_Adaptive_Method_for_Lane_Marking_Detection_Based_on_HSI_Color_Model/links/00b7d51b916b80319f000000.pdf | |||||||
2010 | National University of Ireland | Distance Determination for an Automobile Environment using Inverse Perspective Mapping in OpenCV | IPMに基づく距離推定 | , | https://s3.amazonaws.com/academia.edu.documents/29591239/docualain_issc10.pdf?AWSAccessKeyId=AKIAIWOWYYGZ2Y53UL3A&Expires=1553259605&Signature=HI%2BsbMKexJyAjN38MUUYaOuSoY8%3D&response-content-disposition=inline%3B%20filename%3DDistance_determination_using_inverse_per.pdf | ||||||
2010 | INRIA | General road detection from a single image | Road Boundary Detection | ライン・舗装なし対応。 | 17fps(240x180), | https://www.di.ens.fr/sierra/pdfs/tip10b.pdf | |||||
2010 | Bosch | On-road vehicle detection during dusk and at night | Head Light Detection | State Machine...の[21] | , | https://ieeexplore.ieee.org/document/5548013 | |||||
2010 | University of Zagreb | Real-time Detection and Recognition of Traffic Signs | TSR | 24x24より小さいと検出に失敗 | , | https://www.researchgate.net/publication/224162933_Real-time_detection_and_recognition_of_traffic_signs | |||||
2010 | EPFL | Real-Time Vehicle Tracking for Driving Assistance | Tail Light Detection | 6の[102]、距離推定 | , 5.6% miss | https://infoscience.epfl.ch/record/149699/files/FossatiSF10.pdf | |||||
2010 | National University of Ireland | Rear-lamp vehicle detection and tracking in low-exposure color video for night conditions | Tail Light Detection | color cross-correlation symmetry analysis, Kalman filtering | trackingを援用 | , | https://ieeexplore.ieee.org/abstract/document/5446402/ | ||||
2010 | National University of Ireland | Rear-Lamp Vehicle Detection and Tracking in Low-Exposure Color Video for Night Conditions | Tail Light Detection | 6の[103]、HSVでTL識別 | , | https://ieeexplore.ieee.org/document/5446402 | |||||
2010 | Robust lane markings detection and road geometry computation | Lane Detection,Road Marking Recognition | 解析的モデル、RANSAC | , | http://refbase.cvc.uab.es/files/LSC2010.pdf | ||||||
2010 | Carnegie Mellon University | Stacked Hierarchical Labeling | Road Boundary Detection | KITTIのnon-DL/LiDARでtop(42位) | 7s(>8core), | https://www.ri.cmu.edu/pub_files/2010/9/munoz_eccv_10.pdf | |||||
2010 | State Key Laboratory of Pulsed Power Laser Technology | Taillight detection algorithm based on four thresholds of brightness and color | Tail Light Detection | , | http://en.cnki.com.cn/Article_en/CJFDTOTAL-JSJC201021073.htm | ||||||
2010 | National Chin-Yi University of Technology | Vision-based vehicle detection in the nighttime | Tail Light Detection | , | https://ieeexplore.ieee.org/abstract/document/5533451/ | ||||||
2009 | B-spline modeling of road surfaces with an application to free-space estimation | ステレオによる傾斜推定 | , | https://vision.cs.tum.edu/_media/spezial/bib/wedel_et_al_its09.pdf | |||||||
2009 | Real Time Visual Traffic Lights Recognition Based on Spot Light Detection and Adaptive Traffic Lights Templates | TLR | , | https://www.researchgate.net/profile/Fawzi_Nashashibi/publication/277576745_ArticleFawzi_ITS09_RealTimeVisualTLR/data/556da0a008aec22683061e1d/ArticleFawzi-ITS09-RealTimeVisualTLR.pdf | |||||||
2009 | Traffic Light Detection with Color and Edge Information | TLR | normalized RGBの提案 | , | https://tohoku.repo.nii.ac.jp/?action=repository_action_common_download&item_id=26343&item_no=1&attribute_id=18&file_no=1 | ||||||
2008 | University of Alcal´a | Night Time Vehicle Detection for Driving Assistance LightBeam Controller | Tail Light Detection | , | http://www.robesafe.es/personal/pablo.alcantarilla/papers/Alcantarilla08iv.pdf | ||||||
2008 | Nighttime vehicle detection for intelligent headlight control | Head Light Detection,Tail Light Detection | Real–Adaboost | , 78-97%(head),60-94%(tail) | http://cic.uab.es/Public/Publications/2008/LHB2008/12.pdf | ||||||
2008 | Road Boundary Detection in Complex Urban Environment based on Low-Resolution Vision | Road Boundary Detection | PHT | , | |||||||
2008 | Vehicle Detection at Night Based on Tail-Light Detection | Tail Light Detection | heuristic検出 | , 95.3% | http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.140.4004&rep=rep1&type=pdf | ||||||
2007 | Road-marking analysis for autonomous vehicle guidance | Lane Detection | ルールベース | , | https://www.researchgate.net/profile/Ruediger_Dillmann/publication/221508440_Road-marking_Analysis_for_Autonomous_Vehicle_Guidance/links/0046351c9388f9d211000000/Road-marking-Analysis-for-Autonomous-Vehicle-Guidance.pdf | ||||||
2006 | University of California, San Diego | Video-Based Lane Estimation and Tracking for Driver Assistance: Survey, System, and Evaluation | Lane Detection | , | https://cloudfront.escholarship.org/dist/prd/content/qt1bg5f8qd/qt1bg5f8qd.pdf | ||||||
2005 | Shih-Shinh Huang National Kaohsiung First University of Science and Technology | Driver Assistance System for Lane Detection and Vehicle Recognition with Night Vision | Lane Detection,Tail Light Detection | 上記引用、最速、Vanishing Point Detection158ms | 23ms, 90.51%(40m) | https://www.researchgate.net/profile/Li-Chen_Fu/publication/224623462_Driver_assistance_system_for_lane_detection_and_vehicle_recognition_with_night_vision/links/00b4951702cb342a68000000.pdf | |||||
2019 | Space-Time Slicing Visualizing Object Detector Performance in Driving Video Sequences | Object Detection | 未ラベリングデータでの性能解析を簡易にする可視化 時系列で参考 | ||||||||
2019 | Monocular Plan View Networks for Autonomous Driving | Object Detection | 単眼から鳥瞰図生成 | ||||||||
2019 | Uncertainty Estimation in One-Stage Object Detection | Object Detection | 信頼性定量化 | ||||||||
2019 | Training Object Detectors With Noisy Data | Object Detection | Per-object co-teaching 自動ラベリングするのが目的 ODでラベルノイズの影響を初めて調査 | ||||||||
2019 | Monocular Plan View Networks for Autonomous Driving | Object Detection | BEV,policy生成 入力は640×352にリサイズ Mask RCNN→3D(距離・向き・サイズ)3D推定→再投影→policy | ||||||||
2019 | Monocular 3D Object Detection via Geometric Reasoning on Keypoints | Object Detection | 3D | ||||||||
2019 | Robustness of Object Recognition under Extreme Occlusion in Humans and Computational Models | Object Detection | オクルージョンに対しロバスト性がないことを検証。 Zhangらのvoting modelとSPPの組み合わせの2ステージ。 | ||||||||
2019 | Object Detection Based on the Improved Single Shot MultiBox Detector | Object Detection | SSDの弱点(小物体)解析 Conv4_3 and Conv5_3の受容野が適当な大きさ backboneにfeature fusionを加えGノイズを低減 Conv4_3 and Conv5_3に対し(deconv→)conv→normalize→ReLU→add MAP<1%の微々たる改善 | ||||||||
2019 | RRPN Radar Region Proposal Network for Object Detection in Autonomous Vehicles | Object Detection | radarでRegion proposal Perspective transformation, anchor generation and distance compensationの3ステップ Fatserのようにsize/aspect異なるanchor boxを用意 距離に応じてanchorのスケールを変化 カメラ行列でradarの3D点群を2画像に射影 Points of Interestの概念を導入 RRPNでROIを出力してFasterに渡す | ||||||||
2019 | Precise Synthetic Image and LiDAR (PreSIL) Dataset for Autonomous Vehicle Perception | Object Detection | LiDARデータセット | ||||||||
2019 | Automated Focal Loss for Image based Object Detection | Object Detection | focal lossのパラメータ調整を自動化 | ||||||||
2019 | Complexer-YOLO: Real-Time 3D Object Detection and Tracking on Semantic Point Clouds | Object Detection | Valeo, 3D AVODより精度は低いが、1.5倍程度高速 | ||||||||
2019 | Efficient Incremental Learning for Mobile Object Detection | Object Detection | Samsung+Apple 認識対象を追加していける | ||||||||
2019 | Accurate Monocular 3D Object Detection via Color-Embedded 3D Reconstruction for Autonomous Driving | Object Detection | depth用NNを並列、seg後region,classと統合し3D BB出力 | ||||||||
2019 | GS3D: An Efficient 3D Object Detection Framework for Autonomous Driving | Object Detection | 3D BB orientationを出力する分岐を加える | ||||||||
2019 | The H3D Dataset for Full-Surround 3D Multi-Object Detection and Tracking in Crowded Urban Scenes | Object Detection | Hondaの公開した3D認識用データセット 360 degree LiDAR dataset | ||||||||
2019 | NeurAll Towards a Unified Model for Visual Perception in Automated Driving | Object Detection | detection/segmenattionを単一NNで同時に学習 このようなunified modelは、 Pros:計算効率、学習時間低減 Cons:negative transfer in multi-task learning(人手の部分が誤差に不寛容に;データ設計、アーキテクチャ、ハイパラ、ハードケース) 全タスクに共通のデータセットは少ないので タスク間のバランス等、ⅢD参考 | ||||||||
2019 | TrackNet: Simultaneous Object Detection and Tracking and Its Application in Traffic Video Analysis | Object Detection | 学習は、UA-DETRAC[31] dataset 時系列のBBをtubeと呼ぶ このBBの共通部分の特徴量を抽出するTube Pooling 2つのFCで分類とoffsetを独立に推定 7x7x256の特徴量を256個抽出する lossは5個: crossentropy classification loss (class) regression loss (offset) 上記それぞれtube proposal network前後2個 smooth l1 loss (Fast R-CNNと同じ) | ||||||||
2019 | RetinaMask Learning to predict masks improves state-of-the-art single-shot detection for free | Object Detection | Mask Prediction, Self-Adjusting Smooth L1 Loss | SOTAのRetinaNetの学習方法改良による精度向上により2stage並みの精度に cleaner training dataで精度・収束性向上 Detectronを使用 | RetinaNet | https://arxiv.org/pdf/1901.03353.pdf | https://github.com/facebookresearch/Detectron | ||||
2019 | Mono3D++ Monocular 3D Vehicle Detection with Two-Scale 3D Hypotheses and Task Priors | Object Detection | 単眼からの車両の形状・姿勢推定 ステレオには精度は及ばない | ||||||||
2019 | AdaScale: Towards Real-time Video Object Detection Using Adaptive Scaling | Object Detection | lossが最小になるスケールm_optを出力するよう学習? FPが3割くらい減る効果 高解像度ゆえにFPが多い場合に有効 | ||||||||
2019 | Scale-Aware Trident Networks for Object Detection | Object Detection | 小さい/大きい物体の検出率向上のためにdilation rateの異なるfeatureごとに予測 従来の画像ピラミッドは推論時間に課題 従来の特徴ピラミッド(SSD,FPN)は、スケールごとに表現力(意味情報)が異なることが課題 認識におけるreceptive field単独の影響を初めて論じた dilation rateが大きいと、大サイズの物体の精度が上がる 3ブランチに分け、dilation rate以外のパラメータは共通 各ブランチはFaster R-CNN weight sharingだけでAP~1.5%向上 推論時はブランチを1つにしてもさほど精度は落ちない? ブランチ数は1~4で3がベスト 各ブランチのBBに対しsoft-NMS backboneのResNet-101をDeformableにするだけで2%向上 MXNetで実装 | ||||||||
2018 | Modeling Camera Effects to Improve Visual Learning from Synthetic Data | Object Detection | chromatic aberration, blur, exposure, noise, and color temperatureでDA | ||||||||
2018 | DeepVoting: A Robust and Explainable Deep Network for Semantic Part Detection under Partial Occlusion | Object Detection | 下記引用。 | ||||||||
2018 | Tiny-DSOD Lightweight Object Detection for Resource-Restricted Usages | Object Detection | MobileNet-SSDに対し倍速で精度向上 ただしVGG-SSDよりは精度劣る TITANで105fps(300x300入力) | ||||||||
2018 | Deep continuous fusion for multi-sensor 3D object detection | Object Detection | 3D BB AVODより改善 BEVに変換するContinuous Fusion Layerを介してcameraの複数階層のfmapをLiDARのfmapにfusion | ||||||||
2018 | Pelee A Real-Time Object Detection System on Mobile Devices | Object Detection | SSD+MobileNetに近い、速度は倍。 | ||||||||
2018 | ShuffleDet Real-Time Vehicle Detection Network in On-board Embedded UAV Imagery | Object Detection | SSDの前段をShuffleNetに変えることで、高速化と精度向上を両立 | ||||||||
2018 | Quantization Mimic Towards Very Tiny CNN for Object Detection | Object Detection | 軽量化シート参照 | ||||||||
2018 | Joint Monocular 3D Vehicle Detection and Tracking | Object Detection | Faster R-CNNベースで3D情報でtrackingの探索空間を減らす | ||||||||
2018 | SNIPER: Efficient Multi-Scale Training | Object Detection | Browse state-of-the-artのCOCO,でトップ 画像ピラミッドで全画素扱うのではなく、context regionsのみ 単一解像度の+30%のデータ量で済む 512x512にリサンプリングするため、バッチサイズを20にまで拡大できる 非同期BNが有効 処理画素数をSNIPの1/3に→推論時間もSNIPの1/3 V100で5fps 学習中は物体サイズが既知なので、周辺(context region)のみ処理することで、低解像度画像でも精度を落とさない region proposal networkに基づきnegative instances(chips)を収集 著者MXNet実装 | https://github.com/mahyarnajibi/SNIPER/blob/master/main_train.py | |||||||
2018 | Path Aggregation Network for Instance Segmentation | Object Detection | PANet SNIPERに僅差 Instance Segmentationではトップ Mask R-CNNより大幅に改善 著者Pytorch実 | https://github.com/ShuLiu1993/PANet | |||||||
2018 | RPN-BASED ARCHITECTURE FOR OBJECT DETECTION AND POSE ESTIMATION USING RGB-D DATA | Object Detection | DL姿勢推定の参考、M論 | ||||||||
2018 | Pedestrian Detection based on Deep Fusion Network using Feature Correlation | Object Detection | DSSD+Halfway Fusion+correlation layer RGB-IR2系統入力でNN中央でfusion(concat+NIN) NIN:通常の畳み込み(stride/padding様々、チャンネル増)+1x1conv(チャネル数半減) correlation layerはHadamard prod→sqrt | ||||||||
2018 | CornerNet Detecting Objects as Paired Keypoints | Object Detection | One-stage アンカーベースの課題:大量のアンカーを用意する必要、ハイパーパラメータの多さ RefineDetにメトリックによっては勝ってる 物体の左上を示すヒートマップと右下を示すヒートマップを別のヘッドで出力し、一緒に出力する埋め込みベクトルが同じ物体の時、一致するよう学習 アンカーボックスに比べ高精度かつ実装も簡単で1-stage物体検出の主流になりそう(岡野原氏ツイート) 2-stageのD-RFCN+SNIPには及ばない Heatmaps,Embeddings,offsetを出力 Embedding距離が最小のペアを探索 hourglass(U-NetのskipにRes modile追加?) x2連結 corner pooling(水平方向と垂直方向でpooling) Focal Loss適用 著者Pytorch実装 TITAN Xで240ms/image | https://github.com/princeton-vl/CornerNet | |||||||
2018 | Car Detection using Unmanned Aerial Vehicles: Comparison between Faster R-CNN and YOLOv3 | Object Detection | 航空写真の車検出において、Faster R-CNNよりYOLOv3の方が精度(sensitivity)速度とも優位 Faster R-CNNはTensorflow Object Detection APIを使用。 YOLOv3は著者公開コード(Darknet)使用。 | ||||||||
2018 | Focal Loss for Dense Object Detection | Object Detection | RetinalNet | loss参考、hard negativeの代替 cross entropyにクラス不均衡を補正する重みをかける 下記MegDetからの引用 ICCV2017 student best paper 学習時のクラス不均衡を調整するfocal lossを導入 easy example(正解確率が高い⇔頻度が高い)のcroess entropyのweightを減らす ハイパーパラメータ(α,γ)が増える feature pyramidの各層の出力を、class予測/box予測に分岐し入力 classification subnetの方でfocal lossに使う確率を予測 Online Hard Example Miningより向上 easy exampleも併用すべきということか 画素ごとに領域を抽出 RetinalNetとして引用されている | |||||||
2018 | MegDet: A Large Mini-Batch Object Detector | Object Detection | ECCV2018でトップになったMegviiの(前年の?)論文 COCOでmmAP52.5 mini-batch sizeを256にまで増やすことで高精度化 GPU128台 技術的には、warmup learning rate policyとGPU間BNが新しい | ||||||||
2018 | Recent Advances in Object Detection in the Age of Deep Convolutional Neural Networks | Object Detection | サーベイ 小物体対策、標識認識についてセクションあり Pascal-VOC 画像が小さすぎて過学習を起こしやすい→MS-COCOが通常使われるようになってきた Xu et al., 2017aがトップ(mAP81.2) MS-COCO Deformable R-FCN,Ensembling Mask R-CNNs,がトップ(mAP~50.3。いずれも2017の論文。) | ||||||||
2018 | FSSD: Feature Fusion Single Shot Multibox Detector | Object Detection | SSDのによる高速化 精度はSSD512とDSSD513の中間 1080Tiで35fps PASCAL VOC 2007 mAP=84.5 feature pyramid 2つ作って各レベルから検出を行う | ||||||||
2018 | Single-Shot Refinement Neural Network for Object Detection | Object Detection | RefineDet | Anchor Refinement Module(粗い予測)とObject Detection Module(最終出力)が並列に並び、解像度ごと/間をTransfer Connection Blockで接続 Transfer Connection Block:CRCRCR(ARMとCDMをつなぐ)の真ん中にdeconvをadd ARMは物体の有無だけ、ODMでクラス分類(Hard mining)まで行う。 アスペクト比1:3および3:1のバウンディングボックスをケチる。特徴マップ数を6->4に削減。 論文で他手法との比較が充実 PASCAL VOC2007/2012でトップ Caffe実装公開、Dockerで簡単に実行可能 | https://github.com/daiwc/RefineDet/blob/master | https://qiita.com/nabechi6011/items/5c9f4a3acdf6c55e0a56 | |||||
2018 | Extend the shallow part of Single Shot MultiBox Detector via Convolutional Neural Network | Object Detection | 計算効率優先 SSDの連続する2層の出力を結合して予測に使う特徴マップを増やしている prediction moduleのres layerをconvに置き換え 学習を3段階:SSDのみプレ→追加モジュールのみ学習→全層ファインチューニング | ||||||||
2018 | YOLOv3: An Incremental Improvement | Object Detection | DSSDより高速高精度 24conv+2FC 20convはImageNetで事前学習 7x7sセル単位でBBの5パラとClassの20パラを出力 lossは、BBの中心・サイズ、信頼度(BBの中心があるセル)、不信頼度(BBの中心がないセル)、クラス確率のMSE | https://pjreddie.com/yolo/ | https://www.slideshare.net/ssuser07aa33/introduction-to-yolo-detection-model | ||||||
2017 | Perceptual Generative Adversarial Networks for Small Object Detection | Object Detection | 特徴マップをDに食わせるGAN | ||||||||
2017 | Deformable Convolutional Networks | Object Detection | 下記で引用 convで画素位置のオフセットを算出し、その位置(線形補間)の画素値の加重和としてdeformable convを定義 解説記事 受容野は物体のスケールや形に合わせて適応すべきという考え方に基づく モデルを最初から学習するのではなく、学習済みモデルに deformable の offset field を入れて追加で学習させる 精度がよいのはパラメータ数が増えたからじゃなくて幾何的な変形を捕捉できるようになったから 著者実装(MXNet) | https://github.com/msracver/Deformable-ConvNets | https://qiita.com/keisuke-nakata/items/90f7020f04476b01d07d | ||||||
2017 | An analysis of scale invariance in object detection-snip. | Object Detection | 上記D-RFCN+SNIP Dual path networkをバックボーンとする 分類と検出のデータセットで物体サイズの違いが大きいため、分類用のバックボーンの転移学習がうまくいかない課題に対処 低解像度用に学習されたCNNの方が分類精度は良くなった 高解像度で事前学習+低解像度upsamplingでfine-tuneの方が精度が向上 Deformable-RFCNをベースにする 以下の理由から、End2Endではない Res-Net101をベースにしたもので物体候補を抜き出す Res-Net50をベースにし、localizationの部分を抜いたもので物体の分類を行う GPUのメモリー節約のため 拡大縮小+ランダム抽出(MST)はうまくいかなかった←極端に大きい/小さい物体が入っていたから? →MSTの改良版としてSNIPを提案(解像度によって使うROIサイズ範囲を決めることで、極端な大きさのをはじく) 学習時のメモリを節約する工夫として、ROIクロップにおいて、ROI中の物体数最多のROIを選択→未選択の物体を最も多く含むROIを選択→反復 解説 https://www.slideshare.net/DeepLearningJP2016/dlan-analysis-of-scale-invariance-in-object-detection-snip | ||||||||
2017 | Speed/accuracy trade-offs for modern convolutional object detectors | Object Detection | Faster RCNN, R-FCN, SSDの精度-速度トレードオフの解析 精度なら(Inception-)ResNet特徴抽出のFaster RCNN 速度ならResNet特徴抽出のSSD | ||||||||
2017 | SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving | Object Detection | conv積層→ConvDetでBB・クラス抽出→filtering 車載向け組み込みプロセッサは消費電力の制約(WS250Wに対しNVIDIA Xavierは20W) モデルサイズが小さいと通信オーバーヘッドが減らせる(モデル更新時のみ?) 精度は100%recallにおけるprecisionを高くすべき 漏れが一切なく、過剰検出(FP)を極力減らす Faster R-CNNと違い、RPNをend-to-endで学習 RPNとの違いは、クラスも同時に出力する点 | ||||||||
2017 | DSSD : Deconvolutional Single Shot Detector | Object Detection | SSDの小さい物体(浅い特徴マップしか使っていない)を検出しづらい課題への対策として、深い情報を使う。 SSDの後段にDeconvolutional Moduleを追加 一旦全SSD層を通過させて、その後各スケールを順に抽出しprediction moduleに出力 Deconvolutional Moduleでは、deconv-conv-BNと、対応するスケールのSSD出力にCBRCBをかませた出力の要素積をPrediction Moduleに接続 SSD部前段はVGG16よりResNet101の方がよい ただし素のSSDで同じ交換をしても無意味 Prediction Moduleの改良 クラス・位置ずれ回帰誤差の2系統に入力する前に、res layer(1x1convx3-add)追加 SSDでも効果がある(mAP+0.7%) 1:1.6のDefault boxを追加 | ||||||||
2016 | A unified multi-scale deep convolutional neural network for fast object detection | Object Detection | KITTIのcar/pedistrian detectionでは2016年以前でトップ 0.4s/GPU | ||||||||
2016 | SSD: Single Shot MultiBox Detector | Object Detection | box中に物体が存在する確率を算出し、それに従いboxを調整 様々な解像度の特徴マップを組み合わせることで物体サイズの多様性に対応 従来手法のproposal generationとリサンプリングを排除し、単一のNNに全ての機能を集約 訓練と推定が同一のNN box候補の選択においてはJaccard係数(共通要素数/和集合要素数)とlossを考慮 Jaccard係数>0.5を全て正解とみなす 正解ボックスとの重複が一番大きいものを一つ選択する方法よりこちらの方が精度が高い。 NNの出力は、デフォルトboxの位置・サイズのオフセット、 デフォルトboxとは、中心位置が同じでアスペクトの異なる複数のbox候補群 lossは位置誤差と信頼度の線形結合 GTを基準にして、位置誤差は中心位置の差のHuberと、サイズ比のlogの和 信頼度lossは、Σdefault boxとGTの一致度×log(softmax(各boxに対するNNの信頼度予測)) 特徴マップ(階層)の数だけboxのスケールを用意 各特徴マップに6種のアスペクト比のdefault boxを用意 ハードネガティブマイニング 実際の画像は物体より背景部分の方が多いため、ほとんどのデフォルトボックスがnegativeになる。 対策として、信頼度最大のものだけを使い、正例と負例(背景)が最大でも3:1になるように調整する。 ベースネットワークは、VGG16を以下のように変更。 ・全結合層6,7を畳み込み層に変更 ・プーリング層5の2×2を3×3に、ストライドを2から1に変更 ・a trous algorigthm(wavelet?)によって畳み込みの範囲を疎に広くしている ・Dropout層と全結合層8は削除 異なる層で異なるスケールのボックスを使うことが重要 計算時間の80%はベースネットワークに費やされている 小さい物体の検出には改善の余地がありそう Faster RCNNが1つの特徴マップにさまざまな大きさの候補領域を適用してるのに対し、SSDは特徴マップがマルチスケール。 物体のスケールに依存しにくくなる? | ||||||||
2016 | YOLO9000: Better, Faster, Stronger | Object Detection | YOLOv2 | ||||||||
2019 | Learning 2D to 3D Lifting for Object Detection in 3D for Autonomous Vehicles | Object Detection | GANで単眼3D BB 明確にputperform | https://arxiv.org/pdf/1904.08494.pdf | |||||||
2017 | Perceptual Generative Adversarial Networks for Small Object Detection | Object Detection | 小物体用GAN 超解像特徴マップ生成(conv x5とResNetの和)→DとPerception Branch(BBopx,class出力)に分岐 | ||||||||
2018 | Deep Learning for Generic Object Detection A Survey | Object Detection | NIN, SqueezeNet, MobileNet, ShuffleNet, EffNet, ShiftNetの比較 テンソル分解はやりつくされて性能が飽和しているが、2値化ベースはこれから | ||||||||
2018 | Object Detection with Deep Learning: A Review | Object Detection | 代表的手法ごとに解説 | ||||||||
2019 | Design of Real-time Semantic Segmentation Decoder for Automated Driving | Semantic Segmentation | FCNとの比較しかない 速度はFCNと同等らしい | ||||||||
2019 | Detecting the Unexpected via Image Resynthesis | Semantic Segmentation | 未知物体を不定性で予測 | ||||||||
2019 | Fast-SCNN: Fast Semantic Segmentation Network | Semantic Segmentation | Toshiba,ContextNetの提案者 PSPNetに対しパラメータ数1/600、精度10%悪化 PSPNetに倣い別タスク(分類;ImageNet,お尻をGAP+softmaxに置き換え)でpre-trainingしたが、効果なし! DeepLabは画像を縮小しさらにそれぞれ特徴マップも縮小するが、本手法は特徴マップ縮小のみ 2分岐して、upsampling後とghost blockをaddする構造が新しい 1024 × 2048で123.5fps | ||||||||
2019 | Single Network Panoptic Segmentation for Street Scene Understanding | Semantic Segmentation | panoptic 単一ネットワーク NNの別々の部分間で低レベル特徴を共有し、マルチタスクを同時に学習させることで性能向上 semantic/instanceの出力をheuristicに融合 lossは、RPNのobjectnessのsoftmax、同regression、classのsoftmax、BBのregression、maskのsigmoid、semantocのsoftmax、weight decay ResNet-50+PPM(上記論文参照) seg分岐の出力をnormしてRPN分岐にcatする instance segmentation branchは、Mask R-CNNをベース 画素ごとに label and an instance idが出力、コンフリクトを避ける | ||||||||
2019 | Deep Layer Aggregation | Semantic Segmentation | 様々なタスクに使えるネットワーク構造 2018Joint Monocular 3D Vehicle Detection and TrackingでROI抽出の特徴量算出に使用 階層ごとにUpsampling分岐を増やし逐次連結 比較対象が最新でない? | ||||||||
2019 | Hybrid Task Cascade for Instance Segmentation | Semantic Segmentation | instance segmentationのSOTA SenseTime Mask R-CNNにsemantic segmentation branchを追加 Cascade R-CNNとの違いは、各段階でdetection, mask prediction, semantic segmentationを結合する点 タスク間の関係が考慮される点がベター FishNetを採用 浅い層の情報を後半に連結できる instance/detectionともに、ResNeXtより1.5%程度向上 SoftNMS適用 classwise balance samplingは適用せず 何がうまくいって、何が効果がなかったのかの経験談が参考になる 工数見積もり:reproducing/20days,tuning/30days,expolre new ideas/30days コード公開予定 ECCV2018発表スライドが参考になる http://presentations.cocodataset.org/ECCV18/COCO18-Detect-MMDET.pdf | https://github.com/open-mmlab/mmdetection | |||||||
2019 | Semantic Nighttime Image Segmentation with Synthetic Stylized Data, Gradual Adaptation and Uncertainty-Aware Evaluation | Semantic Segmentation | 夜間に拡張 スタイル変換、徐々に夜間に近づけていく学習 uncertainty-aware intersection-over-union (UIoU) metricを提唱 confidence thresholdではじいたグループが加わった版 RefineNetと比較 | ||||||||
2019 | AuxNet Auxiliary tasks enhanced Semantic Segmentation for Automated Driving | Semantic Segmentation | マルチタスク(seg+depth) encoderはResNet-50で共通 decoderはタスクごとに分岐するが似た構造(deconv+addを3回) 比較対象がSegNetなので大したことない | ||||||||
2019 | UPSNet: A Unified Panoptic Segmentation Network | Semantic Segmentation | FPN(U-Net)の出力をsemantic/instance headに分け、両者の出力をpanoptic headで統合 deformable convolution/Mask R-CNN 組合せと同程度の精度 semantic classはstuffとthingの2つのみ | ||||||||
2019 | Panoptic Feature Pyramid Networks | Semantic Segmentation | FAIR Instance segmentation branchは、Mask R-CNNをFPNで改良 semantic segmentation branchは、FPNの各階層の出力にconv+USでスケールを合わせてaddするだけの非常に単純なもの 1/4resの出力に1x1conv+bilinear upsampling+soft maxで最終出力 BNをchannel-wise affine transformationで置き換え DeeplabV3+に対しmIoU-0.5,FLOPs1.6x | ||||||||
2018 | ThunderNet: A Turbo Unified Network for Real-Time Semantic Segmentation | Semantic Segmentation | 精度犠牲にしてICNetより高速(10ms) | ||||||||
2018 | BiSeNet Bilateral Segmentation Network for Real-time Semantic Segmentation | Semantic Segmentation | ICNetを精度・速度で超える | https://arxiv.org/pdf/1808.00897.pdf | https://github.com/parasdahiya/BiSeNet-Road-Segmentation | ||||||
2018 | A Comparative Study of Real-time Semantic Segmentation for Autonomous Driving | Semantic Segmentation | 下記同著者のCVPR | ||||||||
2018 | RTSeg: Real-time Semantic Segmentation Comparative Study | Semantic Segmentation | 速度に注目したサーベイ Enc:MobileNet, Dec:UNetがベスト精度 | ||||||||
2018 | ContextNet Exploring Context and Detail for Semantic Segmentation in Real-time | Semantic Segmentation | ICNetに劣る(mIoU-3.4%、速度-30%) 下記からの引用 2分岐add(浅いNNでdetail、深いNNでcontext抽出) | ||||||||
2018 | Training of Convolutional Networks on Multiple Heterogeneous Datasets for Street Scene Semantic Segmentation | Semantic Segmentation | 下記論文の著者、下記からの引用 Cityscapes, GTSDB and Mapillary Vistasの3つのデータセットを使った学習は初めて。 3階層ラベル 各階層を別のアーキで推論 各バッチ内でクラスの比率が同じになるよう配分 クラス間imbalanceの原因は、道路や建物など大面積を占めるもの。 データセット間で同一クラスに10^3オーダーの画素数の開き imbalance対策は、以下の記述しかない placing classes with the similar order of examples in the same classifier and thus all classes have bigger probability to be represented in the same batch BBしかラベルがないデータセットを使えるように階層lossを提案 | ||||||||
2018 | ExFuse Enhancing Feature Fusion for Semantic Segmentation | Semantic Segmentation | DeepLab v3+と同等 | ||||||||
2018 | High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs | Semantic Segmentation | LSGANでsemantic labelだけから画像生成 物体境界情報を併用し画質向上 GがE-D2重入れ子のような構造 pix2pixがベース 学習データ生成に使える | ||||||||
2018 | Understanding Convolution for Semantic Segmentation | Semantic Segmentation | decorderにDUC(Dense Upsampling Convolution)、encoderにHDC(Hybrid Dilated Convolution)を適用 DUCはpixel shuffler HDCは連続する3層で異なるdilation rate 同じdilateion rateだと真の受容野もスパースになり、近傍画素間で整合性がなくなる(gridding problem) | ||||||||
2018 | S4-Net: Geometry-Consistent Semi-Supervised Semantic Segmentation | Semantic Segmentation | 少数のRGB-Dフレームからの学習 たった4クラス semi-supervisedのためのgeometric consistency constraints lossが新しい | ||||||||
2018 | Panoptic Segmentation | Semantic Segmentation | FAIR Semantic/Instance Segmentationを融合したタスクの提案 アルゴリズムの提案はない? 定量評価はpanoptic quality (PQ) metricIoUがinstanceごとの和になる点が従来と違う 既存の使えるデータセットは3つだけ(Cityscapes, ADE20k, Mapillary Vistas) | ||||||||
2018 | Context Encoding for Semantic Segmentation | Semantic Segmentation | EncNet Amazon,SenseTime共著 FCNがベース SE的にチャネル重みを出して掛ける。 この途中の出力にFCをかけSE(Semantic Encoding)-lossを算出。 lossが物体の大きさによらない。 global semantic informationの効率的な利用。 SE-lossのGTは正解マップから直接作っている(含まれているクラスを列挙したベクトル?) 同著者のEncoding Layerの手法(2017)をベースにしている Context Encoding:codebookと特徴マップの各画素(1x1xC)の残差を重み調整しReLU+BN→全画素・codebook間平均 Featuremap Attention:上記残差にFC+sigmoidでscaleを算出 adaptive instance normalizationとSE-Netにインスピレーションを得たという記述 特定クラスの強調が目的? PASCAL VOC 2012でDeepLabv3に対し+0.2% ADE20KでPSPNet(ResNet269ベース)に対し-0.29% | http://hangzh.com/ | |||||||
2018 | Searching for Efficient Multi-Scale Architectures for Dense Image Prediction | Semantic Segmentation | DeepLabv3+の著者(Google AI) NAS Google vizier(random search)を使う 現在最良の解周辺を探索 random searchでもそこそこいける EncNetに対しmAP+1.7%、演算量はDeepLabv3+の半分 最後にDense Prediction Cell(DPC)というDAGを接続 5分岐を連結 atrous convのrateを探索 最初にdil. rate1x2のconvを入れるのがベストだった→近接情報がより重要 各候補は30K iter.のみ+early stopping 高速化の工夫として、network backboneを小規模にし、backboneの上に接続するDPCのみ学習し、 explore 28K DPC architectures across 370 GPUs over one week. modified Xception network backbone 1GPUでナイーブ探索だと1週間余り→90min.に短縮 以下のproxy(代理) taskに置き換えることで高速化 最初はMobileNet-v2 backbone(コスト~1/20)で探索 最初に2^5-1通りのアーキ探索空間 proxy/real taskのスコアの相関係数は0.46(不十分に思えるのだが...) 目的関数はmIoU | ||||||||
2018 | End to End Video Segmentation for Driving Lane Detection For Autonomous Car | Semantic Segmentation | GCN+BRは2017Large Kernel Mattersのパクリ(なのに言及せず) 前処理として、学習用画像の空とボンネット(上下)をカット、チェッカーボードを用いた歪曲補正 augmentationは回転 | ||||||||
2018 | ICNet for Real-Time Semantic Segmentation on High-Resolution Images | Semantic Segmentation | mIoUはPSPに劣るが、速度は5倍 階層処理1/4,8,16でのみ学習 Cascade feature fusion unitが特徴() DeepLabv3(ResNet50)に比べ、mIoU-1.8%でx7.2高速化 | https://github.com/hszhao/ICNet https://github.com/aitorzip/Keras-ICNet | |||||||
2018 | Path Aggregation Network for Instance Segmentation | Semantic Segmentation | PANet, InstanceのSOTA Information propagationに着目したMask-R-CNNの改良 detection→things prediction、segmentation→stuff predictionの2つの分岐をマージ | ||||||||
2018 | Mask R-CNN | Semantic Segmentation | detectionとsegmentationを同時に行う ROI-Alignの解説はこちら https://qiita.com/yu4u/items/5cbe9db166a5d72f9eb8 Region proposal(サブ画素ずれあり)を3x3分割(feature mapのサイズにそろえる) grid上のfeat mapの値をbilinear interpolationで求める TF実装ではresizeで高速化されている 3x3のセルごとにave/maxを取る(RoI Poolingと同じ) | ||||||||
2018 | Encoder-decoder with atrous separable convolution for semantic image segmentation | Semantic Segmentation | DeepLab-v3+ ベンチマークでトップ http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?cls=mean&challengeid=11&compid=6&submid=15347 2018/Feb. poolingの代わりにstrideで異スケール Xeptionのdepthwiseで高速化 一旦低解像度にする必要性:バックボーンのResNet101の出力が低解像度なので、より高解像度にするには途中から変更しなければならない encoderで解像度向上 encoderのlow level特徴を入力 超解像の手法が応用できないか? ResNet-101の場合、出力を1/16解像度にするには最後の3ブロック(9層)、1/8にするには26ブロック(78層)もの変更が必要 クラス分類の最後の層のstrideを除外してatrous convに置き換えて、output stride(OS)を32から16(8)に抑えている。 encoderの後半はマルチスケールコンテキスト情報の統合 separable convで演算量40%減 bilinear upasmplingをdecorderに置き換えて精度0.8%向上 decorderはResNet101のConv2を入力して3x3convx2がベスト 1x1convによるチャネル数削減により精度が向上 途切れ・癒着がなくなり定性的にも良好 XecptionのMaxPoolingを、stride=2のseparableに置き換え Cityscapesの評価ではMapillaryが同精度 和訳 https://qiita.com/mine820/items/14e7c556b358dbc4ee9a | https://github.com/tensorflow/models/tree/master/research/deeplab https://github.com/bonlime/keras-deeplab-v3-plus | |||||||
2018 | Evaluating Bayesian Deep Learning Methods for Semantic Segmentation | Semantic Segmentation | posterior推定にMC dropout/Concrete dropoutを使い、MCの方がよかった IOU以外のベイズ推定の評価方法としてを提案 ベースとしてDeepLab-v3+ using Xception posteriorの近似としてのvariational distributionとのKLを最小化 | ||||||||
2018 | Improving Semantic Segmentation via Video Propagation and Label Relaxation | Semantic Segmentation | 動画で学習データ(ラベル)を増やす | ||||||||
2017 | Large Kernel Matters -- Improve Semantic Segmentation by Global Convolutional Network | Semantic Segmentation | localizationのために分類のFCN/GAPを大カーネルconvで代替 受容野を稼ぐために大カーネル 階層処理, Global Convolutional Network(GCN), Boundary Refinement(BR)が特徴 k × k convを1xk+kx1に分解することで、計算量を2/kに抑える 階層処理は、下位GCN出力のdeconv→上位GCN出力へのadd BRは、CRCのRes blockにすぎない 従来はlocalizationに重きを置いていたために分類精度が低い 分類性能は受容野の広さが重要 物体のスケールに依存してしまう ベンチマークではだいぶ下 | ||||||||
2017 | DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs | Semantic Segmentation | atrous(dilated) conv. | ||||||||
2016 | RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation | Semantic Segmentation | 階層入力→階層ごと処理後upsample+sumで統合→Chained residual pooling CRPで解像度が整合取れないように見える点が疑問 著者実装 | https://github.com/guosheng/refinenet | |||||||
2016 | Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning | Semantic Segmentation | MC dropout;Uncertaintyを同時に推定 | ||||||||
2015 | Object Detectors Emerge in Deep Scene CNNs | Semantic Segmentation | 上記からの引用 有効受容野はずっと小さい | ||||||||
2015 | Fully convolutional networks for semantic segmentation | Semantic Segmentation | intersection-over-union (IOU)評価を提唱 | ||||||||
2018 | Deep Ordinal Regression Network for Monocular Depth Estimation | Depth Estimation | KITTIでトップ http://www.cvlibs.net/datasets/kitti/eval_depth.php?benchmark=depth_prediction http://www.robustvision.net/leaderboard.php?benchmark=depth | ||||||||
2018 | GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose | Depth Estimation | SenseTime, CVPR Rigid Structure ReconstructorとNon-rigid Motion Localizerの直列 Rigid...はdepthとposeを独立にCNNで出す Non-rigid...は前者から算出されたOFの整形 forward/backward OFのconsistencyに基づき教師無し学習を行う | ||||||||
2018 | Learning Depth from Monocular Videos using Direct Methods | Depth Estimation | CMU, CVPR depthとposeを独立にCNNで算出しwarped imageを生成、targetと比較 pose CNNにDiffrentiable Direct Visual Odometryを追加する点が新規 失敗例として、歩行者・自転車、輝度飽和、テクスチャレス(幅の広い道路) | ||||||||
2018 | Unsupervised Learning of Depth and Ego-Motion from Monocular Video Using 3D Geometric Constraints | Depth Estimation | Google, CVPR forward/backward depthから点群を作り、ICP loss ペア画像の2つのdepthに基づきそれぞれwarpしクロスで整合性を見る 2つの点群をそれぞれ幾何変換しクロスで整合性を見る | ||||||||
2017 | Unsupervised Learning of Depth and Ego-Motion from Video | Depth Estimation | depth,poseをそれぞれU-Net構造で推定 3フレーム使う 1フレームからdepth推定 3フレーム,depthからpose推定 著者実装 | https://github.com/tinghuiz/SfMLearner | |||||||
2019 | Understanding Neural Architecture Search Techniques | NAS | ENASがランダムサーチを大きく超えない理由の解析 | ||||||||
2019 | Template-Based Automatic Search of Compact Semantic Segmentation Architectures | NAS | セマセグ DeepLab v3+の1/100でmIoU-14% RL | ||||||||
2019 | SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers | NAS | 精度,パラメータ数,メモリ量のmulti objective | ||||||||
2019 | Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours | NAS | MnasNetの5000倍高速化、4hrs以下 kernel sizeを分岐する構造で探索するのではなく、単一パスでsizeを探索 | ||||||||
2019 | Searching for MobileNetV3 | NAS | v2に対しMadd2/3、同latencyで分類精度+3% SSDLite実装でmAP-0.1%で4/3倍高速化(Table 6) セマセグのR-ASPPに実装したが速度-10%程度 h-swishで精度向上 MnasNetベース SENet組み合わせ 非公式Pytorch実装(small versionのみpretrained 公開) | https://github.com/kuan-wang/pytorch-mobilenet-v3 | |||||||
2019 | Searching for a robust neural architecture in four gpu hours | NAS | 最速 Baidu DAGのサンプラーを微分可能にした | ||||||||
2019 | Random Search and Reproducibility for Neural Architecture Search | NAS | ランダムサーチベース 単一のNNで学習された重みを層ごとに割り当て使いまわすので、メモリフットプリントは単一アーキ分のみ ProxylessNASが近い DARTSとENASがSOTAとされている ENASよりも精度は改善 速度は、再現確認ができた条件のDARTSと同程度(DARTS論文で再現されたENASの2倍程度) ENAS等の再現性の問題点を指摘(乱数シード、ドキュメント、stage2:本学習&評価の詳細がない) バッチサイズ小さくするとエポック数を増やしたのと同じ効果が期待できそうだが、勾配がノイジーになり、計算時間が増える 著者PyTorch実装 | https://github.com/liamcli/randomNAS_release | |||||||
2019 | Progressive Differentiable Architecture Search Bridging the Depth Gap between Search and Evaluation | NAS | DARTSをprogressive化 SOTA精度で0.3GPU Days 微分ベースの課題として低精度 原因は探索セル数と評価セル数のgap 探索セル数を段階的に増やしていく Normal cellを1,3,5層 | https://github.com/chenxin061/pdarts | |||||||
2019 | Probabilistic Neural Architecture Search | NAS | ENASと同等の精度・速度 探索空間はDARTSと同じ(normal cell x2/reduction cell交互) DARTSと同じく連続値比率演算混合で微分探索可能にし、empirical Bayes MCで探索 | ||||||||
2019 | NAS-Bench-101 Towards Reproducible Neural Architecture Search | NAS | The dataset contains 423,624 unique neural networks exhaustively generated and evaluated from a fixed graph-based search space. Colab | https://github.com/google-research/nasbench | |||||||
2019 | Multi-Objective Reinforced Evolution in Mobile Neural Architecture Search | NAS | Xiaomi MoreMNAS SR向け PSNR,FLOPs,#paramをobjectiveに 複数objectiveの線形結合ではsub-optimal EAとRLの組合せ mutationをRLでやる Cell-based VDSRとほとんど差無し | ||||||||
2019 | MFAS: Multimodal Fusion Architecture Search | NAS | Fusion NNの構造探索 | ||||||||
2019 | Inductive Transfer for Neural Architecture Optimization | NAS | ENASと同等の精度・速度 | ||||||||
2019 | Fast, Accurate and Lightweight Super-Resolution with Neural Architecture Search | NAS | Xiaomi SR向け SOTAのSelNetを上回る micro(cell)/macro level search cellごとの予測間をLSTMでつなぐ EA | ||||||||
2019 | Evaluating the Search Phase of Neural Architecture Search | NAS | ランダムサーチベース 探索方針はランダムでも先行例の方法でも優位差がないという驚くべき結果 探索空間制約とweight sharing(DARTS,ENAS,NAOに共通)が肝 探索中のランキングは最終的な(スクラッチから学習しなおした)性能と相関が低い | ||||||||
2019 | EAT-NAS Elastic Architecture Transfer for Accelerating Large-scale Neural Architecture Search | NAS | NASNetと全く同等 40GPU days small taskでseedとなる構造を決めてからlarge taskでEAでハイパーパラメータ探索 | ||||||||
2019 | Differentiable Neural Architecture Search via Proximal Iterations | NAS | 0.06GPU daysでNASNet超え Pruning | https://github.com/tanglang96/DDPNAS | |||||||
2019 | Differentiable Architecture Search with Ensemble Gumbel-Softmax | NAS | DARTS改良で精度向上、4->1.5GPU days | ||||||||
2019 | DetNAS: Neural Architecture Search on Object Detection | NAS | 物体検出への応用 20 × 50 models 20 GPUs on 1 day 探索後再学習 evolution algorithm backboneのみの探索 ShuffleNetとの比較:COCOでmAP1-2%向上 後段はFPN/RetinalNet | ||||||||
2019 | Automated Machine Learning | NAS | NASの教科書 | ||||||||
2019 | Adaptive Stochastic Natural Gradient Method for One-Shot Neural Architecture Search | NAS | GAでENASより4倍速い 精度はDARTS 2ndなみ 微分ベース ENAS超え | 0.11GPU days | |||||||
2019 | Accuracy vs. Efficiency Achieving Both through FPGA-Implementation Aware Neural Architecture Search | NAS | latencyをrewardに加える点参考 性能は見るべきところ無し "Then the latency of a task is Kh × Kw ×Tr ×Tc" | ||||||||
2018 | TAPAS Train-less Accuracy Predictor for Architecture Search | NAS | IBM 汎化誤差を予測 | ||||||||
2018 | Taking the Human out of Learning Applications: A Survey on Automated Machine Learning | NAS | サーベイ CIFAR10だと、NASNet-A>ENAS>PNAS>DenseNet meta-learningを広く網羅 | ||||||||
2018 | Searching for Efficient Multi-Scale Architectures for Dense Image Prediction | NAS | semantic segmentationに応用 DeepLab v3+に対しmAP+1.7% 詳細は認識のリスト | ||||||||
2018 | Regularized Evolution for Image Classifier Architecture Search | NAS | Amoeba-Net 精度では僅差でトップ | ||||||||
2018 | Progressive Neural Architecture Search | NAS | RLではなくSMBOを使う モデルアーキテクチャから精度を予測するRNNを用いて, モデルアーキテクチャを全探索する時間を短縮. Keras実装 TF実装 RL,EAはlocal searchの手法 本論文はsequential model-based optimization(SMBO) 構造空間(セルベースの階層構造)をヒューリスティックに、simpleからcomplexに段階的に探索 複数候補(block)の2つの組合せ(cell)を評価し、上位K個に新たにblockを継ぎ足し、を繰り返していく。 5blockのconcatが1cell ShuffleNet/SE-Net<PNAS~NAS |
https://github.com/titu1994/progressive-neural-architecture-search http://github.com/tensorflow/models/ | |||||||
2018 | Neural Architecture Optimization | NAS | 探索空間を連続化 ENASと同等精度・速度(0.3GPU Days) weight sharing(ENASで提案されたparameter sharing)で600倍高速化 string sequence(NN構造)をencoder(LSTM)でパラメータ空間に写像 パラメータ空間で探索 パラメータ空間からNN構造にdecoder()で逆変換 Performance predictor | ||||||||
2018 | MnasNet: Platform-Aware Neural Architecture Search for Mobile | NAS | 高速化のため、latencyを制約に加える | ||||||||
2018 | Efficient Neural Architecture Search via Parameter Sharing | NAS | 2018Faster Discovery of Neural Architectures by Searching for Paths in a Large Modelと同内容 child network間で転移学習 GPU 1台半日程度 全体の計算グラフの一部であるDAGを探索空間とし、全てのエッジに重みを設定し、一部だけを転移学習に使う。 接続の有無も探索対象としている マクロ探索 レイヤー間の接続をDAGとみなし、有効化するエッジと各レイヤのタイプ(以下)を選択。 3x3 convolution 5x5 convolution 3x3 depthwise-separable convolution 5x5 depthwise-separable convolution 3x3 max pooling 3x3 average pooling DAGによってスキップ接続を含めている。 ミクロ探索 CNNの全体像はNASNet同様のものを想定(Normal/Reduction Cellの繰り返し) Convolution(Normal) CellとReduction Cellをそれぞれ別のDAGとする。 入力となる2ノードとそれらに対する処理(以下)を選択。 identity 3x3 separable convolution 5x5 separable convolution 3x3 max pooling 3x3 average pooling マクロ探索とミクロ探索は排他的で、組合せて実行することはできない CIFAR10ではミクロ探索の方が精度が1%よい CutOutだけで0.5%以上向上 conv-BN-ReLUを単位とした 複数層からskipがきた場合にはconcatとした 最後はGAP(FCのパラメータ数削減)+softmax(過学習抑制)とした | https://github.com/melodyguan/enas https://github.com/shibuiwilliam/ENAS-Keras https://github.com/carpedm20/ENAS-pytorch | |||||||
2018 | Dynamic Optimization of Neural Network Structures Using Probabilistic Modeling | NAS | パラメータとハイパーパラメータを1つのlossにまとめてSGDで最適化 lossの勾配はモンテカルロ法で近似 activationの選択のような離散的な選択を連続値変数の確率分布に置き換えることで、SGDのフレームワークに落とし込む | ||||||||
2018 | DARTS Differentiable Architecture Search | NAS | 著者PyTorch実装 候補演算の混合比αを最適化対象とする Ltrainの勾配に基づくパラメータ更新後のLvalを最小化するよう勾配法で探索 ここで用いる擬似学習率の影響が大きい BNには、global moving aveではなくバッチごとの統計を適用 MAC制約あり MobileNet,ShuffleNet,NASNet,AmoebaNet,PNASいずれも500~600M | https://github.com/quark0/darts | |||||||
2018 | AMC: AutoML for Model Compression and Acceleration on Mobile Devices | NAS | 軽量化を参照 | ||||||||
2017 | Peephole: Predicting Network Performance Before Training | NAS | アーキテクチャの性能を予測 LSTM | ||||||||
2017 | Neural Architecture Search with Reinforcement Learning | NAS | NAS CIFAR-10でdepth15/4.2M/5.5%,depth39/37.4M/3.65% DenseNet(depth40/1.0M/5.24%,depth190/25.6M/3.46%)と同等精度 基本2層のLSTM Controllerに複数のchildがぶら下がっている構造のレプリカを並列 childの精度をrewardとする強化学習 現状ではモデルを一から設計する方が特別いいとも悪いともいえない K40 GPUを800台で28日 | ||||||||
2017 | Learning Transferable Architectures for Scalable Image Recognition | NAS | Google,Zoph NASNet CNNに限定 CNNセル(繰り返される構造)の最適化を行う 学習された構造は他タスクに転用可能 Controller RNNとChild Networkの関係はNASのまま サイズ変換によりNormal/Reduction Cellに分類。交互に繰り返す構造。 CIFAR(32x32)用の構造にReduction Cellを足してImageNet(~300x300)用に転用。 各セルの探索内容は、2入力(前の隠れ層)を選択→それらに対する処理を選択→2出力の結合方法を選択 ReLU-conv-BNをセットに Controller RNNを方策勾配法ではなくProximal Policy Optimization(PPO)で学習 PPOは、PGの目的関数を改良したもの。2017年にOpenAIが提唱。 目的関数は、PG:logπA→TRPO:rA→PPO:min(rA, clip(r)A):minは悲観的評価の意味。ただし、r=π/π_old。 DenseNetは上回るが、Shake-Shakeには及ばず SENetの6割のパラメータで全く同精度 DropPath(FractalNetで提唱されている手法)の顕著な効果が見られた。ただし一言書いてあるだけ。 P100を500台で4日 | ||||||||
2017 | Large-Scale Evolution of Image Classifiers | NAS | |||||||||
2017 | Hierarchical Representations for Efficient Architecture Search | NAS | CIFAR-10で3.63% モデル全体を自動設計するのではなく、cellを配置して大まかな構造を与え、cellの中身だけを探索する cellの構造はDenseNet、ただしレイヤの中身は未定 tournament selectionがベース | ||||||||
2017 | Accelerating Neural Architecture Search using Performance Prediction | NAS | SVRによる精度予測 TinyImagenetで25%early stoppingで相関係数0.925 | ||||||||
2019 | GAN-Knowledge Distillation for one-stage Object Detection | Distillation | GAN応用 | ||||||||
2018 | Training Shallow and Thin Networks for Acceleration via Knowledge Distillation with Conditional Adversarial Networks | Distillation | 教師・生徒NNの協調 HintonらがKL-divで生徒に教師の模倣をさせたのに対し、Discを通じての模倣をさせる 3層MLPをDとする 0.3のdropoutをDと生徒に入れる 効果が小さすぎる(ImageNet32でstudent単体で48.2%→47.39%、ちなみにHinton2015だと49.37%) | ||||||||
2018 | KDGAN: Knowledge Distillation with Generative Adversarial Networks | Distillation | GAN応用 MNIST,CIFARでKDでは最高 | https://github.com/xiaojiew1/KDGAN/ | |||||||
2018 | KDGAN Knowledge Distillation with Generative Adversarial Networks | Distillation | GANなら平衡点にたどり着くため応用 student/teacherの出力をDに入力 MNIST,CIFAR-10でDeep model compressionの実験 従来のベストCODISを Naive GANの課題として、データ数とエポック数がたくさん必要、Gの更新で発散・消失がおきやすい distillation loss追加 | https://github.com/xiaojiew1/KDGAN/. | |||||||
2018 | Block-wise Intermediate Representation Training for Model Compression | Distillation | 18Res block単位で教師とのlossを取り、半分の層数のstudentを学習 中間特徴量による蒸留の先行例は、Hint training (i.e., FitNets これをそのまま深いNNに適用しても不安定なので、ブロック単位に変更 student側中間ブロックの入力は教師NNの中間出力 KDより0.3%の向上 | ||||||||
2017 | N2N Learning: Network to Network Compression via Policy Gradient Reinforcement Learning | Distillation | RL応用。 Volvoの人が入ってる。 CIFAR-10/VGG-19で精度-1.1%で35x圧縮 | ||||||||
2017 | Learning Efficient Object Detection Models with Knowledge Distillation | Distillation | Faster-RCNNに適用し、 one-hotでない確率分布(クラス間の関連性)を、教師NNが学習した隠れた情報と解釈 分類のlossは、GTのone-hot(hard target)とのcross entropyおよび教師NNの出力(soft target)とのcross entropyの線形和 logitの確率分布への変換は温度つきsoftmax 2015Hinton+で提案、T大きいほどsoft(正解でないクラスの値が増加)な分布 cross entropyは、背景クラスのみ重み1.5に増やす。 teacherの精度とstudentの高速性(高々1.5x)を両立 解説記事 http://ai.deepx.co.jp/2018/09/25/%E8%92%B8%E7%95%99-%E7%AC%AC2%E5%9B%9E/ 課題1:クラス間インバランス soft target loss(xentropy)において、バックグラウンドクラスのみ1.5倍の重み 課題2:回帰タスクへの蒸留の応用 ナイーブな対応(lossを連続値用に替える)では、教師NNの誤差に制限がないために、それに引っ張られるリスクがある。 そこで、smooth L1 Loss(student vs. GT)に加え、teacherの誤差-マージンよりもstudentの誤差が大きい場合にのみL2 lossを加える。 課題3:知識の継承をより効果的にする 中間層のマッチングを取る。 チャネル数の違いは、studentの出力に1x1 convを加えること(adaptation layer)で解決。 課題4:高速化 studentへの入力を低解像度にする。 高解像度版に近い精度が得られている。 | ||||||||
2017 | Data Distillation: Towards Omni-Supervised Learning | Distillation | 入力に様々な幾何変換を加えて学習済みモデルを適用した出力を教師とする Faster-RCNNに適用し、 種々の幾何変換に対し出力されたboxの和集合を使う 全カテゴリに共通の閾値を使うとバイアスが発生するため、ラベル有無データでannotated instancesの数の平均が一致するよう閾値を選択 | ||||||||
2019 | URNet : User-Resizable Residual Networks with Conditional Gating Module | Model Compression | Samsung 可変圧縮率 組み込みでニーズがあるらしい Scale parameter SでResNet moduleの1分岐にGatingをかける Sによる精度悪化は小さい CGM(conditional gating module)追加で精度向上? | ||||||||
2019 | sharpDARTS Faster and More Accurate Differentiable Architecture Search | Model Compression | DARTSより高精度・軽量・高速探索 0.8 GPU days Sharp Separable Convolutionの導入 PyTorch実装コード記載 Cosine Power Annealing LR Cosine Annealingでは初期にLRがほとんど減らないため改善にimbaranceが生じる問題を改善 | ||||||||
2019 | ResNet Can Be Pruned 60× Introducing Network Purification and Unused Path Removal (P-RM) after Weight Pruning | Model Compression | VGGで0.34%精度落ちでx44圧縮 ResNet-18では0.31%精度落ちでx55圧縮 ADMMが最適解でないことを初めて示し、続くPurification and Unused Path Removalにより圧縮率を2~10倍向上 | ||||||||
2019 | Really should we pruning after model be totally trained? Pruning based on a small amount of training | Model Compression | 学習時に圧縮 8x-9x compression for VGG-19 テスト・学習とも10x高速化 few epoch pretraining->pruning+trainの反復 | ||||||||
2019 | PruneTrain Gradual Structured Pruning from Scratch for Faster Neural Network Training | Model Compression | スパース化しながらスクラッチから学習 FLOPs減らすと急激に悪化 30-50%減らせる 学習時間は20-30%低減 pruningはUnstructured/Structuredに大別 Unstructuredは不規則なデータアクセスパタンのため高速化には不利 Group lasso regularizorをlossに追加 | ||||||||
2019 | Model Compression by Entropy Penalized Reparameterization | Model Compression | Google VGGで1/590サイズ(Top1 3.4%落ち) endto-end で重みの確率分布を考慮し性能と圧縮率を最大化 | ||||||||
2019 | Learning Topological Representation for Networks via Hierarchical Sampling | Model Compression | Hierarchical SamplingによってNNを小規模のNNの連結に再帰的に分解していく 階層構造に分解し、Node embeddingsとして表現 NASに応用できるとインパクトが大きいかもしれない | ||||||||
2019 | How Compact?: Assessing Compactness of Representations through Layer-Wise Pruning | Model Compression | Samsung 精度を向上させ64%パラメータ数減 | ||||||||
2019 | FSNet: Compression of Deep Convolutional Neural Networks by Filter Summary | Model Compression | SSDを1/4,7%落ちに kernelを生成するFilter Summaryを学習 | ||||||||
2019 | Efficient Hybrid Network Architectures for Extremely Quantized Neural Networks Enabling Intelligence at the Edge | Model Compression | ハイブリッドでVGG19メモリ圧縮21.8x ただし精度は67.2→62.1 NNの頭とお尻とres skipのconvのみFP、あとはbinary | ||||||||
2019 | Effective Network Compression Using Simulation-Guided Iterative Pruning | Model Compression | Deep Compressionとの違いは、 percentile threshold reducedではなくoriginal NNに重み更新をして重要度を算出 Deep Compressionに対する精度向上は微々たるもの Deep Compressionは1/100辺りで精度がピークになる(圧縮によって精度向上!) | ||||||||
2019 | DSConv: Efficient Convolution Operator | Model Compression | メモリ1/14、速度10x | ||||||||
2019 | Deep Neural Network Approximation for Custom Hardware: Where We’ve Been, Where We’re Going | Model Compression | サーベイ、ハード実装との関係 | https://arxiv.org/pdf/1901.06955.pdf | |||||||
2019 | Creating Lightweight Object Detectors with Model Compression for Deployment on Edge Devices | Model Compression | SSDのbackboneのResNet-50のchannel pruning model size=16.3MB, FLOPS=2.31G, and mAP=71.2, MobileNet BBより良好 channel pruning→knowledge transfer→channel deletion→knowledge distillation | ||||||||
2019 | CGaP: Continuous Growth and Pruning for Efficient Deep Learning | Model Compression | VGG-19 on CIFAR-100で+0.37%,78.9%pruned,acc 1.4× reinforcing important learning unitsでユニット増やしてからpruning | ||||||||
2019 | Automated Pruning for Deep Neural Network Compression | Model Compression | Deep Compressionより3割程度の圧縮率向上 微分可能にし、学習中にpruningを実行 FCの圧縮が主要(最大~30x)、convはよくて3割程度減 ReLUとsigmoidを組み合わせたpruning functionを導入 3値化のステップを滑らかにした これをweight decayに組み込む? | ||||||||
2019 | Adversarially Trained Model Compression When Robustness Meets Efficiency | Model Compression | adversarial attackに対するロバスト性を両立するモデル圧縮 MobileNetV2と同等のパラメータ数で、adv有無ともに精度1%以上向上 上記からの引用 intrinsic relationship between CNN weight sparsity and adversarial robustnessを調べた ある程度までは、スパースにしたほうがロバスト性が向上 5%以下のスパース性は脆弱 | ||||||||
2018 | Weightless: Lossy weight encoding for deep neural network compression | Model Compression | スパースで誤差を許容するほど、Deep Compressionより圧縮率が高くなる Bloom filterを一般化したBloomier filterを適用 | ||||||||
2018 | Universal Deep Neural Network Compression | Model Compression | Samsung ランダムディザ付加→要素ごと等間隔量子化→fine-tune→lossless coding ディザ付加の効果は、高圧縮率で精度が下がる領域で降下を遅らせること Deep Compressionと大差なし | ||||||||
2018 | Targeted Dropout | Model Compression | Hintonのチーム Dropoutは同一層内のユニット間の相互情報量を最大化し、学習後のスパース化にロバストにする効果がある アプリオリに重要でないと考えられる層にdropoutを入れて学習してからスパース化する方法をtargeted dropoutと呼ぶ 著者コード github.com/for-ai/TD 精度を落とさないpruningは60%まで可能(ResNet-32 on CIFAR-10) dropoutを入れて学習することでpruning rateを上げられるという重要な知見 weightのdropout候補をパーセンタイル(targeting proportion;γ)で決め、それらに対してあるdrop probability;αでdrpout実行 γ=0.75,α=0.66がベスト | ||||||||
2018 | Studying the Plasticity in Deep Convolutional Neural Networks using Random Pruning | Model Compression | filter pruningの効果は、filterを選ぶ基準によるのではなく、DNNの可塑性によるものと主張 random pruningでも他の手法と同等であることを示した(pruning後のfine-tuning前提) 最初から少ないフィルタ数で学習するのとは違うのか? differential pruning(層ごとのフィルタ数に比例した削除率)で0.5%向上;あまり効果なし 4xで15%程度も精度が悪化している pruning手法間比較は充実している random以外ではl1-Norm[27]がベスト いまいちクオリティが低い論文 | ||||||||
2018 | Slim and Efficient Neural Network Design for Resource-Constrained SAR Target Recognition | Model Compression | モデル圧縮の先行例の参考にもなる 重みが閾値以下なら切り捨てるnetwork pruningの比を決める | ||||||||
2018 | ShuffleNet An Extremely Efficient Convolutional Neural Network for Mobile Devices | Model Compression | MobileNetよりパラメータ数を減らし精度向上(ほぼ同等だが) オプショナルなSE block追加で精度向上 1x1GroupConv->channel shuffle->3x3Depthwise->1x1GroupConv->Res add cuda-convnetでサポートされているrandom sparse convolutionがchannel shuffleに相当 pointwise group convが新しい。弊害対策としてchannel shuffle。 1*1 pointwise convolutionは頻出であるが, これはチャネル方向への参照範囲が大きくコストが高い. これをgrouped convolutionに変更し, 更に出力チャネルが全ての入力チャネルを参照できるよう 「channel shuffle」(チャネルの入れ替え)を行う層を提案. | ||||||||
2018 | Rethinking the value of network pruning | Model Compression | 従来のpruning後fine-tuneと(pruning後)ランダム初期化+再学習とが大差なし(むしろ向上)という驚くべき結果 モデル(FLOPs)圧縮分エポック数を増やすと精度向上、一方fine-tuneでエポック数増やしても効果なし pruningはNASの一種という解釈 generalization errorの上界を圧縮前後のweightの相互情報量で理論的に与えた empirical riskの増加を超えるだけgeneralization errorを減らすよう圧縮すれば、 population riskが改善する population riskとは? モデル圧縮は過学習を抑制する正則化として働くという解釈 代表的手法の再実装を含めたPyTorch実装公開 Faster-RCNNへの転移学習 prune後の構造で(クラス分類の)プレ学習をする方がベター Yang+(2017)のFaster-RCNNのPytorch実装を利用 https://github.com/jwyang/faster-rcnn.pytorch 解説 https://qiita.com/f0o0o/items/56f66dc109fb78af320f | https://github.com/Eric-mingjie/Rethinking-network-pruning | |||||||
2018 | Resource-Scalable CNN Synthesis for IoT Applications | Model Compression | 対象クラスを減らすことで学習済みモデルを縮小 class-imbalance issueに対しては、学習におけるクラスごとのインスタンス数をそろえて対処 GoogLeNetで速度3.5x | ||||||||
2018 | Recovering from Random Pruning: On the Plasticity of Deep Convolutional Neural Networks | Model Compression | weightのpruningの手法間比較 | ||||||||
2018 | Quantization Mimic Towards Very Tiny CNN for Object Detection | Model Compression | 蒸留+量子化 教師NNを量子化した上で、量子化された生徒NNを学習 低次元多様体(生徒NN)を高次元の離散データ(教師NNの量子化されたfmap)にfitするのならより簡単 (注)lossの計算時にだけ量子化するのであって、量子化されたweightを学習するのではない lossはR-FCN, Faster R-CNNと同じ INQを引用;weightを2の冪かに制約 INQと違いuniform quantizationにした(丸める値を等間隔にする) ROI poolingでは大きい値を正確に出力することが重要なため VGG-1-4 with R-FCNで、サイズ1/16、速度3x、ハードケースほど精度向上 VGG-1-32は1/600、10xで、medium,hardが元より向上 | ||||||||
2018 | Pruning neural networks is it time to nip it in the bud? | Model Compression | アーキテクチャが重要という主張はRethinking...論文に通じる(ちゃんと引用している) 時間の制約下では、pruningよりも小サイズのNNをscratchから学習する方がよい Fisher-pruning (Theis et al., 2018)を適用 pruningによるlossの増加の解析的近似を導出 | ||||||||
2018 | Pruning at a Glance: A Structured Class-Blind Pruning Technique for Model Compression | Model Compression | 全層を一括で考慮すること(Blindness)とweightではなくフィルタ単位でpruningすること(Structured Pruning)との組合せが新しい ResNet-110を精度を落とさずに53%削除 LeNet-5は97.4%削減、Deep Compressionと同等 sorted L1 normのパーセンタイルで閾値処理という点は普通 | ||||||||
2018 | PocketFlow An Automated Framework for Compressing and Accelerating Deep Neural Networks | Model Compression | ハイパラ最適化+代表的な手法の圧縮+追加学習をひとまとめにしたTFパッケージを公開 | https://github.com/Tencent/PocketFlow | |||||||
2018 | Pelee A Real-Time Object Detection System on Mobile Devices | Model Compression | DenseNetの軽量化 | ||||||||
2018 | NISP: Pruning Networks using Neuron Importance Score Propagation | Model Compression | Rethinking...の引用 最終層の1個前の層において影響の大きいニューロンを特定(Inf-FS [34]適用)して、逆伝搬によってネットワーク内各チャネルの寄与率を推定 | ||||||||
2018 | Network compression using Correlation Analysis of Layer Responses | Model Compression | Rethinking...の引用 | ||||||||
2018 | Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters | Model Compression | KL div.の制約により圧縮率を制御 誤差 vs. 圧縮率のトレードオフ(Pareto-frontier)がDeep Compression, Bayesian Compressionより改善 encoding distribution p and a parameterized variational family qφの間のKL div.をペナルティとして課す 著者TF実装 | https://github.com/cambridge-mlg/miracle | |||||||
2018 | Heterogeneous Bitwidth Binarization in Convolutional Neural Networks | Model Compression | 1-bit,2-bit一様よりも、weightによって1/2-bitを切り替えることで精度向上 | ||||||||
2018 | Heterogeneous Bitwidth Binarization in Convolutional Neural Networks | Model Compression | 1,2,3-bit混在(平均1.4bit) MobileNetにおいて、full precisionに対し-5% 全部3-bitに近い精度 BinaryConnectでは、hard sigmoid(台形)で得られる確率で2値化 ビット数を増やすと、精度が大きく向上(Hubara et al. (2016) and Zhou et al. (2016)) ●N-bit binarization:Tang et al. (2017) ±(残差)平均値で2値化後の残差をさらに2値化、というプロセスを繰り返す 演算量の低減は64/m/n(m,nは入力,重みのビット数) 高速化には入力の2値化も必要 FPGA,ASIC実装の評価? 量子化による汎化誤差低減を相互情報量で説明 | ||||||||
2018 | FLOPs as a Direct Optimization Objective for Learning Sparse Neural Networks | Model Compression | lossはweightにベルヌーイ乱数でマスクをかけて、対数尤度とFLOPsのlossの和を取る 第1項はreparameterization trickで最適化、第2項はREINFORCEを適用 conv層のFLOPs loss算出は、(KwKhCin + 1)(Iw − Kw + Pw + 1)(Ih − Kh + Ph + 1)||z||0 https://github.com/AMLab-Amsterdam/L0_regularization のlossにFLOPsを加えたらしい。 上記コードの手法に対し、同精度でFLOPs3/8 | ||||||||
2018 | Fast On-the-fly Retraining-free Sparsification of Convolutional Neural Networks | Model Compression | On-the-flyでretrainingなしで圧縮 モデルによって、パーセンタイルと上層ほどスパースにするのと優位性が入れ替わる Deep Compressionとの違いは、iteratively pruning and retrainingがないこと | ||||||||
2018 | Exploring Weight Symmetry in Deep Neural Networks | Model Compression | 転置対称性?の制約 channelwise-triangularがベスト。~20%のパラメータ数削減。 VGGで約半分。精度-0.08。 | ||||||||
2018 | EffNet: An Efficient Structure for Convolutional Neural Networks | Model Compression | MobileNet, ShuffleNetと比較 depthwise convolution2層をconvで挟む形のブロック データ圧縮は上層に回す方がいい | ||||||||
2018 | Differentiable Fine-grained Quantization for Deep Neural Network Compression | Model Compression | 不均一ビット数 binary並みの圧縮率(1/30)と、8-bitに近い(が劣る)精度 精度は明らかに落ちてるのでいまいち(CIFAR-10,VGG-16で-1.7%) bi-level optimizationとして定式化 モデルサイズとval.誤差の2項 最適化はDARTSに倣う(Laglangeの未定乗数法?) | ||||||||
2018 | Demystifying Neural Network Filter Pruning | Model Compression | L1ベースのpruningを、フィルタの機能の観点で解析 retrainingは、誤って消したフィルタの機能を残ったフィルタに付加する効果があることが判明 出力のフィルタに対する敏感度のテスト画像間平均を寄与率としてpruningを決定する方法により、精度劣化が大幅に低減(詳細な数値が示されていない) | ||||||||
2018 | Deep Neural Network Compression by In-Parallel Pruning-Quantization | Model Compression | CLIP-Q 容量をVGG16で1.4%にまで削減し精度は微増。 Deep Compressionの半分くらいで、精度は向上 ThiNetは精度が低下 MobileNet,ShuffleNetも1/7程度に軽量化 pruningと量子化を同時に。 weightをパーセンタイルでクリップし、内側は全て0→外側を決められたビット数で分割→各分割Gr.ごとに平均値に置き換え ガウス過程前提のベイズ最適化で層ごとに独立にパーセンタイルとビット数を決定 Bayesian optimizationは非凸で微分困難なブラックボックスの目的関数の最小化に使える。 ガウス過程とは、観測点以外の平均値と分散をガウス関数で補間推定するための仮定。 平均値の大きさ(期待値)と分散の大きさ(未探索領域)をバランスよく探索する方針。 理論は2014Bayesian Optimization with Inequality Constraintsに依拠 objectiveのpriorをガウシアンで仮定し(?)たときの平均と共分散から算出されるexpected improvementを最大化する探索を行う BO参考 http://krasserm.github.io/2018/03/21/bayesian-optimization/ 実装の詳細が'We implemented CLIP-Q in Caffe and used the public Bayesian optimization libraries of [61], [62].'としか書かれていないので再現困難。 高速化は別途工夫が必要 | ||||||||
2018 | Deep Learning is not a Matter of Depth but of Good Training | Model Compression | LRのcos減衰の周期を整数倍で増やしていく Stochastic Gradient Descent with Warm Restarts (SGDR) LRの範囲は1E-1~-6 ResNet-110を1/10の層数にして同等の精度を得ている | ||||||||
2018 | DCFNet Deep Neural Network with Decomposed Convolutional Filters | Model Compression | フィルタを既知の基底に分解 パラメータ数削減と正則化 | ||||||||
2018 | Compressing Neural Networks using the Variational Information Bottleneck | Model Compression | VGG-16で60%程度のメモリフットプリント、モデルサイズはNetwork-Slimmingより数割改善 variational approximationで求めた層内mutual informationの最小化による冗長性低減 著者実装 | https://github.com/zhuchen03/VIBNet | |||||||
2018 | ADMM-NN An Algorithm-Hardware Co-Design Framework of DNNs Using Alternating Direction Method of Multipliers | Model Compression | pruningとquantizationの組合せにより231xのモデル圧縮(AlexNet) AlexNetでpruning率13xで精度劣化無し pruning率が55%を超えないと速度は向上しない?(break-even weight pruning ratio; Figure 4) pruningは不規則なスパース性とインデクスのためにオーバーヘッドが生じる quantizationでは影響が重畳するため誤差が拡大する Deep compressionはquantizationではなくclusteringを使っている 2段階(pruning率決定→pruning→quantizationをbreak-evenに基づき確定するステップをはさみ2回) break-even ratioを超える(効果がある)層だけ選択的にpruning率を上げていく ADMMにより、AlexNetでpruningは72hrs、quantizationは24hrs Deep Compressionの3.6x圧縮 Caffe実装 | http://bit.ly/2M0V7DO | |||||||
2018 | Adam Induces Implicit Weight Sparsity in Rectifier Neural Networks | Model Compression | 東芝。解析方法参考。 ReLU+L2 loss+Adamでgroup sparsityが現れることを証明。 Adamだと70% sparsity(weight vectorのL2 normのヒストを示す)。momentam SGDでは全くスパース性は現れない。 CIFAR-10に対し、4層MLPで60%削減しても精度落ちず。 | ||||||||
2018 | A Survey on Methods and Theories of Quantized Neural Networks | Model Compression | 量子化のサーベイ | ||||||||
2017 | xUnit: Learning a Spatial Activation Function for Efficient Image Restoration | Model Compression | DnCNNのパラメータを46%削減して同等性能 | ||||||||
2017 | Towards Accurate Binary Convolutional Neural Network | Model Compression | |||||||||
2017 | Soft Weight-Sharing for Neural Network Compression | Model Compression | 上記サーベイの引用(parameter sharingの代表として) minimum description length (MDL) problemを変分法で解いている。 1. Nowlan & Hinton (1992)の手法に従い、プレ学習済み重みの範囲を等分する複数のガウシアン(θでパラメトライズ)で初期化 2. error cost(log(p(X|w)))とcomplexity cost(priorとposteriorの近似間のKL距離)の和をlossとしたwとθの学習 3. 各成分の平均に置き換え(量子化) ResNetで45x圧縮だが2%誤差増大 | ||||||||
2017 | On Compressing Deep Models by Low Rank and Sparse Decomposition | Model Compression | 高速性よりもメモリ削減(圧縮率)に主眼 weight sharing , quantization, and Hoffman codingとの組み合わせも可能(示唆のみ) 重み行列を低ランク行列とスパース行列の和に分解 Greedy Bilateral Decompositionを提案 圧縮率0.55まで精度はほぼ落ちない 上記2015Learning both...と比較。圧縮率0.55まではほぼ同等。 | ||||||||
2017 | MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications | Model Compression | depth wise conv(separable filter) | ||||||||
2017 | Learning Efficient Convolutional Networks through Network Slimming | Model Compression | channel-level sparsityを課す 40/60%pruningで精度向上 | ||||||||
2017 | Incremental Network Quantization Towards Lossless CNNs with Low-Precision Weights | Model Compression | quantizationのみでAlexNet 53x圧縮(Deep Comp.は35x) 精度が向上する(4,5bitでref.超え。ImageNet/ResNet50でerr-1.59%) 絶対値が大きいweightのみ量子化→量子化されてないweightのみ再学習、のループ 量子化するweightの選択をランダムにしてもそれほど悪化しない 5->3 bitで53x→89x圧縮率向上 低ビット化により、乗算をFPGA上でビットシフトに置き換えられ高速化につながる 著者Caffe実装 | https://github.com/AojunZhou/Incremental-Network-Quantization | |||||||
2017 | Channel Pruning for Accelerating Very Deep Neural Networks | Model Compression | 下記AMCは本手法とRLの組合せ。 フィルタ係数にスパース性を課してLASSOを解く。 1層ずつ逐次的に決定。 Caffe実装 VGG16のpretrained model単独でも公開されているので使える。 | https://github.com/yihui-he/channel-pruning | |||||||
2017 | Bayesian Compression for Deep Learning | Model Compression | VGGで最大95x圧縮(精度-0.2%)、51x高速化(バッチサイズ256) dropoutによる変分ベイズ推定を用いてパラメータに分布を導入. 導入した分布によりスパースなパラメータの学習が誘引される. evidence-lower-bound (ELBO)をlossとする 対数事後確率 著者PyTorch実装 | https://github.com/KarenUllrich/Tutorial_BayesianCompressionForDL | |||||||
2017 | Analytical Guarantees on Numerical Precision of Deep Neural Networks | Model Compression | 量子化による誤差の理論上限 8bitから急激に悪化 weight precision(Bw)をactivation precision(Ba)より大きくすると誤差が大きく低下 Bw=Ba,Bw=Ba+3を実験 | ||||||||
2017 | A Survey of Model Compression and Acceleration for Deep Neural Networks | Model Compression | |||||||||
2016 | XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks | Model Compression | パラメータとアクティベーションを二値化した場合にはXNOR+popcountで高速化できる | ||||||||
2016 | SqueezeNet AlexNet-level accuracy with 50x fewer parameters and 0.5MB model size | Model Compression | convのタップ削減、ビット数削減でAlexNetのパラメータ数を最大1/510に削減 | ||||||||
2016 | EIE: Efficient Inference Engine on Compressed Deep Neural Network | Model Compression | weight sharingによって疎な積演算に | ||||||||
2016 | EIE: Efficient Inference Engine on Compressed Deep Neural Network | Model Compression | 上記のハード実装偏 DRAM→SRAMでエネルギー効率120x GPU上で13x高速化 pruningによるsparsityは4-25% quantizationは4bit compressed sparse column (CSC) format | ||||||||
2016 | Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding | Model Compression | Pruning-quantization-Huffman encodingの3段階 pruningと再学習の反復 k-meansでクラスタリングし平均値に置き換える量子化(weight sharing) 勾配も量子化 圧縮率は、ビット数、接続数、クラスタ数で決定される AlexNetで35x、精度変わらず 上記サーベイで'state-of-art performance among all parameter quantization based methods' | ||||||||
2016 | Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications | Model Compression | Samsung kernelをvariational Bayesian matrix factorizationで決めたランクでTucker分解 圧縮率はnetworkに強く依存 AlexNetでx5.46、VGG16でx1.09 | ||||||||
2015 | Learning both Weights and Connections for Efficient Neural Networks | Model Compression | 上記の引用 解説記事https://qiita.com/naomi7325/items/701a74e65cd504ae26a9 Deep compressionのpruning部分は同じらしい AlexNetで1/9、VGG-16で1/13にパラメータ数を圧縮 閾値以下の重みは0に置き換え 入力/出力が0のニューロンは削除 その後、再学習 学習において、L2正則化が有効 上位層のpruningの方が精度劣化が少ない 下位層は広範囲に伝播するから当然か convとFCで一方を固定し一方だけをpruning pruning後初期化してしまわないこと pruningと再学習のセットを反復することで、5xの圧縮率を9xにまで増やせた。 Pytorch実装 | https://github.com/jack-willturner/DeepCompression-PyTorch | |||||||
2015 | FitNets Hints for Thin Deep Nets | Model Compression | |||||||||
2019 | Optimizing CNN-based Object Detection Algorithms on Embedded FPGA Platforms | Acceleration | FPGA実装参考 | ||||||||
2019 | MetaPruning: Meta Learning for Automatic Neural Network Channel Pruning | Acceleration | AMCと同等 | ||||||||
2018 | Stacked Filters Stationary Flow For Hardware-Oriented Acceleration Of Deep Convolutional Neural Networks | Acceleration | stacked filters stationary flow: | ||||||||
2018 | NetAdapt Platform-Aware Neural Network Adaptation for Mobile Applications | Acceleration | latency,消費電力をmetricにする AMCに類似 同等のlatencyでMobileNetV2以上の精度 | ||||||||
2018 | Learning to Prune Filters in Convolutional Neural Networks | Acceleration | |||||||||
2018 | Hybrid Pruning Thinner Sparse Networks for Fast Inference on Edge Devices | Acceleration | チャネル→weight pruningの2段階 ResNet50で精度-1.7%、パラメータ数73%減 weightのL1ノルムでソート 精度を満たすかチェックしながら間引きを増やしていく 層ごとにsparsityを適応的に決める 閾値を、weight絶対値のmean+std×scaleとする | ||||||||
2018 | GANAX A Unified MIMD-SIMD Acceleration for Generative Adversarial Networks | Acceleration | 同著者 | ||||||||
2018 | FlexiGAN An End-to-End Solution for FPGA Acceleration of Generative Adversarial Networks | Acceleration | transposed convolutionのFPGAによる高速化 | ||||||||
2018 | Flexible Deep Neural Network Processing | Acceleration | 2倍程度高速化 | ||||||||
2018 | Differentiable Training for Hardware Efficient LightNNs | Acceleration | FPGA実装、精度-0.3%で30x高速化 weightを2のべきに限定 先行例と同様、lossの微分はstraight-through estimator (STE)で計算 | ||||||||
2018 | B-DCGAN:Evaluation of Binarized DCGAN for FPGA | Acceleration | 全要素の2値化 | ||||||||
2018 | AMC: AutoML for Model Compression and Acceleration on Mobile Devices | Acceleration | アーキテクチャ固定なのでNASではない?精度のみならずlatencyをターゲットにできる。よりシンプルな方法。 従来人手で決められていた特徴マップのスパース性を自動的に抽出 各層のパラメータ(入出力チャネル数,画像サイズ,stride,k,FLOPs,)をagentが決める agentが決めた圧縮率に従い、2015Learning both weights...のchannel pruningを使う Faster R-CNN4倍高速化でmAP+0.1% TensorFlow Lite frameworkで時間評価 Intel® Distribution of OpenVINO™ toolkit fp32→int8で1.5x前後高速化 | ||||||||
2018 | A Survey on Acceleration of Deep Convolutional Neural Networks | Acceleration | サーベイ | ||||||||
2017 | Solving internal covariate shift in deep learning with linked neurons | Acceleration | |||||||||
2017 | IDEAL: Image DEnoising AcceLerator | Acceleration | BM3D,DJDD高速化 | ||||||||
2017 | FFT-Based Deep Learning Deployment in Embedded Systems | Acceleration | FFTによる学習・推論の高速化 | ||||||||
2017 | Extremely Large Minibatch SGD: Training ResNet-50 on ImageNet in 15 Minutes | Acceleration | 1024GPU,15min.,74.9%(精度犠牲なのが難あり) | ||||||||
2017 | Accurate, Large Minibatch SGD Training ImageNet in 1 Hour | Acceleration | Facebook,GPU8->256台で29倍高速化、バッチサイズ8kにすることで大規模並列化を可能にした バッチサイズを増やしても精度が劣化しないテクニックとして2つ提案 linear scaling:バッチサイズに比例して学習率を増やす バッチサイズと学習率は交換可能な概念 バッチサイズを増やすとバッチ内平均のために個々のサンプルによる勾配の比率が減るため、学習率を減らすのと等価(という理解で正しい?) warmup:最初5エポックで徐々に学習率を増やす Sharp minima問題(質の悪い局所解)はバッチサイズが大きいだけでなく学習率が低い場合にも起こる 256GPU,1 hr,76.3 % | ||||||||
2017 | A New Approach to Compute CNNs for Extremely Large Images | Acceleration | 画像のみならずモデルも分割して並列処理 画像分割すると(バッチ)正規化が問題。この解決方法が肝。 サブ画像間の重複領域は入力画像の画素値を使ってパディング; dictionary padding BSP (bulk synchronization parallel) model VGGのReLU1-4_1をstyle loss、ReLU4_2をcontent lossに使う | ||||||||
2016 | From High-Level Deep Neural Models to FPGAs | Acceleration | DNN Weaver開発者の論文 FPGAとGPUの比較 GTX 650Tiに対し数倍~数百倍の高速化 | ||||||||
2014 | Speeding up convolutional neural networks with low rank expansions. | Acceleration | conv filterの近似 |