Computer Vision – ACCV 2020 Workshops: 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 – December 4, 2020, Revised Selected Papers ... Vision, Pattern Recognition, and Graphics)

This book constitutes the refereed post-conference proceedings of four workshops held at the 15th Asian Conference on Computer Vision, ACCV 2020, which was held in Kyoto, Japan, in November/ December 2020.* The 13 papers were carefully reviewed and selected from the following two workshops: Machine Learning and Computing for Visual Semantic Analysis (MLCSA) and Multi-Visual-Modality Human Activity Understanding (MMHAU). *The conference and workshops were held virtually. Preface 6 Organization 7 Contents 8 Machine Learning and Computing for Visual Semantic Analysis (MLCSA) 10 Spatial and Channel Attention Modulated Network for Medical Image Segmentation 11 1 Introduction 11 2 Related Work 13 2.1 Medical Image Segmentation 13 2.2 Deep Attention Network 14 3 Spatial and Channel Modulate Network 14 3.1 The Mainstream of the Encoder-Decoder Architecture 15 3.2 Spatial Attention Module (SAM) 16 3.3 Channel Attention Module (CAM) 17 3.4 Spatial and Channel Attention Module (SCAM) 18 4 Experimental Setup and Results 19 4.1 Database 19 4.2 Evaluation Results 20 5 Conclusion 22 References 23 Parallel-Connected Residual Channel Attention Network for Remote Sensing Image Super-Resolution 26 1 Introduction 26 2 Method 28 2.1 Network Architecture 28 2.2 Parallel Connection Module 30 2.3 Residual Channel Attention Block 31 2.4 Loss Function 31 3 Experiments 32 3.1 Data Preparation and Parameter Settings 32 3.2 Quantitative Results 32 3.3 Visual Results 33 4 Conclusions 35 References 36 Unsupervised Multispectral and Hyperspectral Image Fusion with Deep Spatial and Spectral Priors 39 1 Introduction 40 2 Related Work 42 2.1 Traditional Methods 42 2.2 Deep Learning Based Methods 43 3 Proposed Method 43 3.1 Problem Formulation 44 3.2 The Proposed Deep Spatial and Spectral Priors (DSSP) 44 4 Experiment Result 46 5 Conclusion 51 References 51 G-GCSN: Global Graph Convolution Shrinkage Network for Emotion Perception from Gait 54 1 Introduction 54 2 Related Work 55 3 Background 56 4 Proposed Method 57 4.1 Global Graph Convolution 58 4.2 Shrinkage Block 60 5 Experiments 61 5.1 Datasets and Implementation 61 5.2 Ablation Experiments 62 5.3 Comparison with State-of-the-Art Methods 62 6 Conclusions 63 References 64 Cell Detection and Segmentation in Microscopy Images with Improved Mask R-CNN 66 1 Introduction 66 2 Related Work 68 2.1 Cell Detection Approaches 69 2.2 Cell Segmentation 69 3 Method 70 3.1 Weight-Selection Strategy 71 3.2 Focal-Loss Based Mask R-CNN 73 4 Experiments Results 73 4.1 Dataset 73 4.2 Evaluation Metric 74 4.3 Performance Comparison 74 5 Conclusion 76 References 77 BdSL36: A Dataset for Bangladeshi Sign Letters Recognition 79 1 Introduction 79 2 Related Works 82 3 Our BdSL Dataset 83 3.1 Data Collection and Annotation 83 3.2 Dataset Split 85 3.3 Comparison with Other Datasets 86 4 Experimental Evaluation 87 4.1 Experiment Settings 87 4.2 Evaluation Metrics 88 4.3 Classification and Detection with Deep Learning Networks 88 4.4 Further Analysis 89 5 Conclusion 92 References 92 3D Semantic Segmentation for Large-Scale Scene Understanding 95 1 Introduction 95 2 Related Works 97 2.1 Range Image-Based Methods 97 2.2 Voxel-Based Methods 97 2.3 Point-Based Methods 98 3 3D Point Cloud Semantic Segmentation Network 98 3.1 Data Preparation 99 3.2 Data Loading 99 3.3 Background for Model Building 100 3.4 Model Building 101 3.5 Conditional Random Field 102 4 Experiments and Results 102 4.1 Implementation Details 102 4.2 Datasets 102 4.3 Evaluation 103 4.4 Results and Discussion 103 4.5 Time and Space Complexity 104 4.6 Comparison with Other Datasets 105 4.7 Proposed Network with CRF 106 4.8 Ablation Study 107 5 Conclusions and Future Work 108 References 108 A Weakly Supervised Convolutional Network for Change Segmentation and Classification 111 1 Introduction 111 2 Related Work 113 2.1 Fully Supervised Change Detection 113 2.2 Weakly Supervised and Unsupervised Change Detection 114 3 Proposed Method 115 3.1 W-CDNet Overview 115 3.2 Siamese Network 115 3.3 Comparison Block 115 3.4 CSC Module 116 3.5 Model Variations 118 3.6 Training 118 4 Experiments and Results 119 4.1 Datasets 119 4.2 Performance Measures 120 4.3 Influence of Residual and Comparison Blocks on Performance 121 4.4 Influence of CRF Refinement on Performance 122 4.5 Comparison Results on AICD Dataset 123 4.6 Comparison Results on HRSCD Dataset 123 5 Conclusion 124 References 125 Multi-Visual-Modality Human Activity Understanding (MMHAU) 128 Visible and Thermal Camera-Based Jaywalking Estimation Using a Hierarchical Deep Learning Framework 129 1 Introduction 129 2 Related Work 131 3 Algorithm 132 3.1 Classification Step 133 3.2 Semantic Segmentation Step 134 3.3 Hierarchical Framework: Training 134 3.4 Hierarchical Framework: Testing 135 4 Experimental Results 135 4.1 Comparative Analysis 137 5 Summary 139 References 140 Towards Locality Similarity Preserving to 3D Human Pose Estimation 142 1 Introduction 142 2 Related Work 145 2.1 Example-based Approaches 145 2.2 Locality Similarity Preserving for Human Pose Estimation 145 3 Background 146 4 Proposed Model 147 4.1 Human Pose Split Using Kinematic Priors 147 4.2 The Locality Similarity Preserving for Human Pose 147 5 Experiment 149 5.1 Datasets 150 5.2 Comparison Approaches 150 5.3 Evaluation Protocols 151 5.4 Quantitative Evaluation on Human3.6M 151 5.5 Quantitative Evaluation on HumanEva-I 153 5.6 Qualitative Evaluation on MPII 154 5.7 Ablation Study 154 6 Conclusion 156 References 156 Iterative Self-distillation for Precise Facial Landmark Localization 160 1 Introduction 160 2 Related Works 161 2.1 Facial Landmark Localization 161 2.2 Self-distillation 162 3 Method 162 3.1 Loss Terms 163 3.2 Self-distillation (SD) 164 4 Result 165 4.1 Dataset 165 4.2 Model 166 4.3 Implementation Detail 167 4.4 Experimental Result 167 4.5 Conclusion 171 References 171 Multiview Similarity Learning for Robust Visual Clustering 174 1 Introduction 174 2 Related Works 177 2.1 Graph Learning Revisited 177 2.2 Self-representation via Low-Rank Constraint 178 3 The Proposed Framework of RMvSL 178 3.1 Numerical Algorithm 180 4 Experiments 183 4.1 Datasets 183 4.2 Experimental Results and Analysis 184 4.3 Parameters Sensitivity and Convergence 186 5 Conclusions 188 References 188 Real-Time Spatio-Temporal Action Localization via Learning Motion Representation 190 1 Introduction 190 2 Related Work 192 3 Learning Motion Representation 193 3.1 MENet 194 3.2 Long Range Motion Information 194 3.3 Short Range Motion Information 195 4 Experiments 197 4.1 Datasets and Metrics 197 4.2 Training Settings 197 4.3 Ablation Study 198 4.4 Comparison to the State of the Art 199 5 Conclusion 201 References 201 Author Index 205