Computer Vision – ACCV 2020: 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 – December 4, 2020, Revised Selected Papers, Part II (Lecture Notes in Computer Science)

The six volume set of LNCS 12622-12627 constitutes the proceedings of the 15th Asian Conference on Computer Vision, ACCV 2020, held in Kyoto, Japan, in November/ December 2020.* The total of 254 contributions was carefully reviewed and selected from 768 submissions during two rounds of reviewing and improvement. The papers focus on the following topics: Part I: 3D computer vision; segmentation and grouping Part II: low-level vision, image processing; motion and tracking Part III: recognition and detection; optimization, statistical methods, and learning; robot vision Part IV: deep learning for computer vision, generative models for computer vision Part V: face, pose, action, and gesture; video analysis and event recognition; biomedical image analysis Part VI: applications of computer vision; vision for X; datasets and performance analysis *The conference was held virtually. Preface Organization Contents – Part II Low-Level Vision, Image Processing Image Inpainting with Onion Convolutions 1 Introduction 2 Related Work 2.1 Traditional Non-learning Methods 2.2 Learning-Based Methods 3 Approach 3.1 Onion Convolution 3.2 Discussion on Implementation 3.3 The Network Architecture 3.4 Loss Functions 4 Experiments 4.1 Qualitative Comparison 4.2 Quantitative Comparison 4.3 Ablation Study 5 Conclusion References Accurate and Efficient Single Image Super-Resolution with Matrix Channel Attention Network 1 Introduction 2 Related Work 3 Matrix Channel Attention Network 3.1 Network Structure 3.2 Matrix in Matrix 3.3 Matrix Channel Attention Cell 3.4 Hierarchical Feature Fusion 3.5 Comparison with Recent Models 4 Experimental Results 4.1 Datasets and Evaluation Metrics 4.2 Implementation Details 4.3 Comparisons with State-of-the-Art Algorithms 4.4 Ablation Study 5 Conclusion References Second-Order Camera-Aware Color Transformation for Cross-Domain Person Re-identification 1 Introduction 2 Related Works 2.1 Supervised Person ReID 2.2 Unsupervised Domain Adaptation 2.3 Unsupervised Person ReID 3 Proposed Method 3.1 Color Statistics Calculation 3.2 Mean and Covariance Matching 3.3 Transformation for Different Cameras 3.4 Progressive Unsupervised Learning 4 Experiments 4.1 Implementation Details 4.2 Ablation Study 4.3 Comparison with State-of-the-Arts 5 Conclusions References CS-MCNet: A Video Compressive Sensing Reconstruction Network with Interpretable Motion Compensation 1 Introduction 1.1 Related Works 1.2 Contribution 2 Methodology 2.1 Priliminary Reconstruction Module 2.2 Multi-hypothesis Motion Compensation Module 2.3 Residual Reconstruction Module 2.4 Learning Algorithm 3 Experiment 3.1 Implementation Details 3.2 Comparison with the State-of-the-Art 3.3 Reconstruction Under Noise 3.4 Discussion 4 Conclusion References MCGKT-Net: Multi-level Context Gating Knowledge Transfer Network for Single Image Deraining 1 Introduction 2 Related Work 2.1 Single Image Deraining 2.2 Multi-level Learning Network 3 The Proposed MCGKT-Net 3.1 The Mainstream of the Encoder-Decoder Architecture 3.2 Knowledge Transfer Module 3.3 Multi-level Context Gating Module 4 Experimental Results 4.1 Experimental Setting-Up 4.2 Comparison to the State-of-the-art Methods 4.3 Ablation Studies 5 Conclusion References Degradation Model Learning for Real-World Single Image Super-Resolution 1 Introduction 2 Related Work 2.1 Single Image Super-Resolution 2.2 Real-World SISR 2.3 Degradation Model Learning 3 The Proposed Method 3.1 Formulation of Image Degradation Model 3.2 Degradation Model Learning 3.3 SISR Model Learning 4 Experimental Results 4.1 Experiment Setup 4.2 Datasets and Implementation Details 4.3 Ablation Study 4.4 Experiments on Real-World SISR 5 Conclusions References Chromatic Aberration Correction Using Cross-Channel Prior in Shearlet Domain 1 Introduction 2 Cross-Channel Prior in Shearlet Domain 2.1 LCA Correction Model Using CC-SD Prior 2.2 Optimization Using ADMM 3 PSF Estimation Based on Wave Propagation 3.1 Wave Propagation Model 3.2 Establish the LUT Between Wd, W040 and R0 4 Experiments 4.1 Experiment Image Sets and Parameters Setting 4.2 Comparison of Simulation Results 4.3 Comparison of Real-Captured Image Results 5 Conclusion References Raw-Guided Enhancing Reprocess of Low-Light Image via Deep Exposure Adjustment 1 Introduction 2 Related Work 2.1 Traditional Methods 2.2 Learning-Based Methods 2.3 Datasets 3 RAW-Guiding Exposure Time Adjustment Network 3.1 Motivation 3.2 RAW Image Pipeline 3.3 Network Architecture 4 Experiments 4.1 Training Details 4.2 Comparison to Traditional Methods 4.3 Comparison to Learning-Based Methods 4.4 Ablation Study 5 Conclusion References Robust High Dynamic Range (HDR) Imaging with Complex Motion and Parallax 1 Introduction 2 Related Work 2.1 Motion Handling Methods in HDR Reconstruction 2.2 Multi-camera HDR Reconstruction 2.3 Deep Registration Networks 3 Proposed Method 3.1 Approach Overview 3.2 Pyramidal Alignment 3.3 Merging Network 3.4 Loss Function 4 Experiments 4.1 Datasets 4.2 Implementation Details 4.3 Analysis of Single-Camera Case 4.4 Analysis of Multi-camera Case 5 Ablation Study 6 Conclusion References Low-Light Color Imaging via Dual Camera Acquisition 1 Introduction 2 Related Work 3 Method 3.1 Framework 3.2 RefEC Net 3.3 RefColor Net 3.4 RefSR Net 4 Experiments 4.1 Implementation 4.2 Comparison 5 Conclusion References Frequency Attention Network: Blind Noise Removal for Real Images 1 Introduction 2 Related Work 3 Frequency Attention Network 3.1 Network Architecture 3.2 Wavelet Transform 3.3 Spatial-Channel Attention Block (SCAB) 4 Experiments 4.1 Implement Details 4.2 Ablation Study 4.3 Experiments on Synthetic Noise 4.4 Experiments on Real-World Noise 5 Conclusion References Restoring Spatially-Heterogeneous Distortions Using Mixture of Experts Network 1 Introduction 2 Related Work 2.1 Image Distortion Restoration 2.2 Multiple Image Distortion Restoration 2.3 Multi-task Learning 3 Spatially-Heterogeneous Distortion Dataset 4 Our Method 4.1 Model Overview 4.2 Mixture of Parameter Shared Experts 4.3 Attentive Feature Fusion 5 Experiments 5.1 Implementation Details 5.2 Comparison to the Other Methods 5.3 Model Analysis 6 Conclusion References Color Enhancement Using Global Parameters and Local Features Learning 1 Introduction 2 Related Work 3 Proposed Method 3.1 Overview 3.2 Global Parameters Extractor Subnetwork 3.3 Local Features Extractor Subnetwork 3.4 Loss Function 4 Experiments 4.1 Training Details 4.2 Ablation Study 4.3 Alternative Architecture 4.4 Comparison with the State-of-the-Art Methods 5 Conclusion References An Efficient Group Feature Fusion Residual Network for Image Super-Resolution 1 Introduction 2 Related Work 2.1 Lightweight Neural Network 2.2 Recent Residual Blocks 3 Proposed Method 3.1 Network Architecture of GFFRN 3.2 Lightweight GFFRN (GFFRN-L) 3.3 Group Feature Fusion Residual Block (GFFRB) 4 Experimental Results 4.1 Datasets 4.2 Implementation and Training Details 4.3 Comparisons with State-of-the-Art Methods 4.4 Discussion 5 Conclusions and Future Works References Adversarial Image Composition with Auxiliary Illumination 1 Introduction 2 Related Work 2.1 Image Composition and Harmonization 2.2 Generative Adversarial Networks 3 Proposed Method 3.1 Local Harmonization 3.2 Global Harmonization 3.3 Adversarial Training 4 Experiment 4.1 Dataset 4.2 Experiment Setting 4.3 Experiment Analysis 5 Conclusion References Overwater Image Dehazing via Cycle-Consistent Generative Adversarial Network 1 Introduction 2 Related Work 2.1 Methods Based on Hand-Crafted Features 2.2 Methods Based on CNN Features 2.3 Image Dehazing Dataset 3 Proposed Method 3.1 Generator 3.2 Discriminator 3.3 Loss Function 3.4 Dehazed Image Quality Assessment 4 Experiment 4.1 Dataset 4.2 Experimental Settings 4.3 Qualitative Results on Real Images 4.4 Qualitative and Quantitative Results on Synthetic Images 4.5 Dehazed Images Ranking 5 Analysis and Discussion 5.1 Effect of Resize Convolution 5.2 Effect of Perceptual Loss 6 Conclusion References Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning 1 Introduction 2 Related Work 3 Proposed Model 3.1 Motivation and Overview 3.2 Overall Architecture 3.3 Attentive Auxiliary Feature Block 4 Experiments 4.1 Dataset and Evaluation Metric 4.2 Implementation Details 4.3 Ablation Study 4.4 Comparison with State-of-the-Art Methods 4.5 Running Time and GFLOPS 4.6 Perceptual Metric 5 Conclusion References Multi-scale Attentive Residual Dense Network for Single Image Rain Removal 1 Introduction 2 Background and Related Work 3 Proposed Method 3.1 Design of MARD-Net 3.2 Multi-scale Attention Residual Block (MARB) 3.3 Loss Function 4 Experiments 4.1 Datasets and Performance Metrics 4.2 Training Details 4.3 Evaluation on Synthetic Datasets 4.4 Evaluations on Real Rainy Images 4.5 Ablation Studies 5 Conclusion References FAN: Feature Adaptation Network for Surveillance Face Recognition and Normalization 1 Introduction 2 Related Work 3 Feature Adaptation Network 3.1 Framework Overview 3.2 Disentangled Feature Learning 3.3 Paired and Unpaired Feature Adaptation 4 Experiments 4.1 Implementation Details 4.2 Ablation Study 4.3 Comparisons with SOTA Methods 5 Conclusions References Human Motion Deblurring Using Localized Body Prior 1 Introduction 2 Related Works 3 Human and Scene Deblurring 3.1 Generator 3.2 Discriminator 4 Optimization 5 Experiment Procedure 6 Experimental Results 7 Conclusion References Synergistic Saliency and Depth Prediction for RGB-D Saliency Detection 1 Introduction 2 Related Work 3 Method 3.1 Prediction Module: The First Stage 3.2 Prediction Module: The Second Stage 3.3 Discriminator 3.4 Complete Training Loss 4 Experiments 4.1 Datasets and Evaluation Metrics 4.2 Implementation Details 4.3 Comparison with State-of-the-Art Methods 4.4 Ablation Study 5 Conclusions References Deep Snapshot HDR Imaging Using Multi-exposure Color Filter Array 1 Introduction 2 Related Work 3 Proposed Deep Snapshot HDR Imaging 3.1 Framework Overview 3.2 Luminance Estimation 3.3 Luminance-Normalized HDR Image Estimation 3.4 Network Architecture 4 Experimental Results 4.1 Setups 4.2 Validation Study of Our Framework 4.3 Comparison with Other Methods 4.4 Comparison with State-of-the-Art Methods Using a Single or Multiple LDR Images 4.5 Limitation 5 Conclusion References Deep Priors Inside an Unrolled and Adaptive Deconvolution Model 1 Introduction 2 Related Works 3 Unrolled Deconvolution Network 3.1 Adaptive Deconvolution Module 3.2 Learning Image Priors 4 Experiments 4.1 Non-blind Deconvolution on Synthesized Datasets 4.2 Noise Robustness 4.3 Convergence Efficiency and Runtime Comparison 5 Conclusions References Motion and Tracking Adaptive Spatio-Temporal Regularized Correlation Filters for UAV-Based Tracking 1 Introduction 2 Related Work 2.1 Spatial Regularization 2.2 Spatio-temporal Regularization 3 ASTR-CF 3.1 Objective Function of ASTR-CF 3.2 Optimization of ASTR-CF 3.3 Target Localization 4 Experimental Results 4.1 Quantitative Evaluation 4.2 Ablation Study 5 Conclusion References Goal-GAN: Multimodal Trajectory Prediction Based on Goal Position Estimation 1 Introduction 2 Related Work 3 Problem Definition 4 Goal-GAN 4.1 Motion Encoder (ME) 4.2 Goal Module (GM) 4.3 Routing Module (RM) 4.4 Generative Adversarial Training 4.5 Losses 5 Experimental Evaluation 5.1 Benchmark Results 5.2 Assessing Multimodality of Predictions on Synthetic Dataset 5.3 Qualitative Evaluation 6 Conclusion References Self-supervised Sparse to Dense Motion Segmentation 1 Introduction 2 Related Work 2.1 Motion Segmentation 2.2 Sparse to Dense Labeling 3 Proposed Self-supervised Learning Framework 3.1 Annotation Generation 3.2 Deep Learning Model for Sparse to Dense Segmentation 4 Experiments 4.1 Datasets and Implementation Details 4.2 Sparse Trajectory Motion-Model 4.3 Dense Segmentation of Moving Objects 5 Conclusion References Recursive Bayesian Filtering for Multiple Human Pose Tracking from Multiple Cameras 1 Introduction 2 Related Work 3 Method 3.1 Prediction Step 3.2 Update Step 3.3 Initialization Step 3.4 Inference 4 Experiments 4.1 Quantitative Comparison 4.2 Tracking 4.3 Ablation 4.4 Parameters 4.5 Runtime Analysis 5 Conclusion References Adversarial Refinement Network for Human Motion Prediction 1 Introduction 2 Related Work 2.1 Human Motion Prediction 2.2 Prediction Refinement 2.3 Adversarial Learning 3 Methodology 3.1 Overall Framework 3.2 Refinement Network 3.3 Adversarial Learning Enhanced Refinement Network 3.4 Training Loss 4 Experiments 4.1 Datasets and Evaluation Metrics 4.2 Implementation Details 4.3 Quantitative Comparisons 4.4 Qualitative Visualizations 5 Ablation Studies 5.1 Different Components in Our ARNet 5.2 Multi-stage Analysis 6 Conclusions References Semantic Synthesis of Pedestrian Locomotion 1 Introduction 1.1 Related Work 2 Methodology 2.1 States and Actions 2.2 Semantic Trajectory Policy Network (STPN) 2.3 Human Locomotion Network (HLN) 2.4 Reward Signal 2.5 Policy Training 3 Experiments 3.1 Datasets 3.2 Training and Evaluation Details 3.3 Results 4 Conclusions References Betrayed by Motion: Camouflaged Object Discovery via Motion Segmentation 1 Introduction 2 Related Work 3 A Video Registration and Segmentation Network 3.1 Motion Representation 3.2 Differentiable Registration Module 3.3 Motion Segmentation Module 4 MoCA: A New Moving Camouflaged Animal Dataset 4.1 Detailed Statistics 5 Experiments 5.1 Datasets 5.2 Baselines 5.3 Training and Architecture Details 6 Results 6.1 Results on the MoCA Benchmark 6.2 Results on DAVIS2016 Benchmark 7 Conclusions References Visual Tracking by TridentAlign and Context Embedding 1 Introduction 2 Related Work 3 Proposed Method 3.1 Region Proposal with Scale Adaptive TridentAlign 3.2 Classification with Context-Embedded Features 3.3 TridentAlign and Context Embedding Tracker 4 Experiments 4.1 Implementation Details 4.2 Quantitative Evaluation 4.3 Ablation Study 5 Conclusion References Leveraging Tacit Information Embedded in CNN Layers for Visual Tracking 1 Introduction 2 Method 2.1 Discriminative Correlation Filter Tracker 2.2 Incorporating Information of CNN Layers 2.3 Regularizing the Filter 2.4 Implementation Details 3 Experiments 3.1 The Effects of Regularization on Single Layer 3.2 Employing Multiple Layers of CNN 3.3 Scale Adaptation 3.4 Activation vs. Style 3.5 Preliminary Analysis 3.6 Comparison with State-of-the-Art 4 Conclusion References A Two-Stage Minimum Cost Multicut Approach to Self-supervised Multiple Person Tracking 1 Introduction 2 Related Work 3 AutoEncoder-Based Multicut Approach 3.1 Multicut Formulation 3.2 Deep Convolutional AutoEncoder 3.3 AutoEncoder-Based Affinity Measure 4 Experiments and Results 4.1 Ablation Study 4.2 Results 5 Conclusion References Learning Local Feature Descriptors for Multiple Object Tracking 1 Introduction 2 Related Work 2.1 Embedding Learning: Person Re-ID 2.2 Embedding Learning: Local Feature Descriptors 3 ODESA-Based Tracker 3.1 Object Detector 3.2 ODESA Embedding 3.3 Data Association Stage 4 Results 5 Discussion 5.1 ODESA Object Representation 5.2 Sensitivity to Occlusions, Background and Detection Noise 6 Conclusions References VAN: Versatile Affinity Network for End-to-End Online Multi-object Tracking 1 Introduction 2 Related Work 3 Proposed Method 3.1 Overview 3.2 Candidate Refinement 3.3 Versatile Affinity Network 3.4 Event-Aware Training 3.5 Candidate Association 4 Experiments 4.1 Implementation Details 4.2 Evaluation on MOT Benchmarks 4.3 Ablation Study 5 Conclusions References COMET: Context-Aware IoU-Guided Network for Small Object Tracking 1 Introduction 2 Related Work 2.1 Generic Object Tracking on Surveillance Videos 2.2 Detection/Tracking of Small Objects from Aerial View 3 Our Approach 3.1 Offline Proposal Generation Strategy 3.2 Multitask Two-Stream Network 4 Empirical Evaluation 4.1 Implementation Details 4.2 Ablation Analysis 4.3 State-of-the-Art Comparison 5 Conclusion References Adversarial Semi-supervised Multi-domain Tracking 1 Introduction 2 Related Works 3 Proposed Tracker 3.1 Private-Shared Multi-domain Learning 3.2 Network Architecture 3.3 Adversarial Representation Learning 3.4 Self-supervised Representation Learning 3.5 Semi-supervised MDL Training 4 Online Tracking 4.1 Tracking-by-Detection 4.2 Stochastic Universal Negative Mining 4.3 Implementation Details 5 Experiment 6 Conclusion References Tracking-by-Trackers with a Distilled and Reinforced Model 1 Introduction 2 Related Work 3 Methodology 3.1 Preliminaries 3.2 Visual Tracking as an MDP 3.3 Learning Tracking from Teachers 3.4 Student Architecture 3.5 Tracking After Learning 4 Experimental Results 4.1 Experimental Setup 4.2 Results 5 Conclusions References Motion Prediction Using Temporal Inception Module 1 Introduction 2 Related Work 3 Methodology 3.1 Temporal Inception Module 3.2 Graph Convolutional Network 3.3 Implementation and Training Details 4 Evaluation 4.1 Datasets 4.2 Results 4.3 Ablation Study 5 Conclusion and Future Work References A Sparse Gaussian Approach to Region-Based 6DoF Object Tracking 1 Introduction 2 Related Work 3 Probabilistic Model 3.1 Preliminaries 3.2 General Formulation 3.3 Discrete Scale-Space Formulation 4 Optimization 4.1 Gaussian Equivalence 4.2 Regularized Newton Method 4.3 Gradient and Hessian Approximation 5 Implementation 6 Evaluation 6.1 Experiments 6.2 Results 6.3 Essential Aspects 7 Conclusion References Modeling Cross-Modal Interaction in a Multi-detector, Multi-modal Tracking Framework 1 Introduction 2 Related Work 2.1 Multi-object Tracking 2.2 Cross-Modal Attention Mechanism 3 Approach 3.1 Pipeline Overview 3.2 Multi-detector Proposal Collection 3.3 Cross-Modal Attention Module 3.4 Linear Programming Formulation 4 Experiment 4.1 Dataset 4.2 Metrics 4.3 Implementation Detail 4.4 Ablation Study 4.5 Analysis on Multiple Detectors Performance 4.6 Benchmark Evaluation 5 Conclusion References Dense Pixel-Wise Micro-motion Estimation of Object Surface by Using Low Dimensional Embedding of Laser Speckle Pattern 1 Introduction 2 Related Work 3 Overview of the Method 3.1 Principle 3.2 System Configuration and Algorithm 4 Implementation 4.1 Representation of Object Surface 4.2 Embedding Speckle Pattern for Each Pixel 4.3 Making the Embedded Vectors Spatially Consistent 4.4 Optimizing Surface Parameters 5 Experiments 5.1 Evaluating the Calculated Displacement with Respect to the Real Offset 5.2 Comparison with Accelerometers 5.3 Visualizing Various Movements 6 Conclusion References Author Index