PRICAI 2021: Trends in Artificial Intelligence: 18th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2021, Hanoi, Vietnam, ... Part III (Lecture Notes in Computer Science)

This three-volume set, LNAI 13031, LNAI 13032, and LNAI 13033 constitutes the thoroughly refereed proceedings of the 18th Pacific Rim Conference on Artificial Intelligence, PRICAI 2021, held in Hanoi, Vietnam, in November 2021. The 93 full papers and 28 short papers presented in these volumes were carefully reviewed and selected from 382 submissions. PRICAI covers a wide range of topics in the areas of social and economic importance for countries in the Pacific Rim: artificial intelligence, machine learning, natural language processing, knowledge representation and reasoning, planning and scheduling, computer vision, distributed artificial intelligence, search methodologies, etc. Part III includes two thematic blocks: Reinforcement Learning, followed by Vision and Perception. Preface Organization Contents – Part III Reinforcement Learning Consistency Regularization for Ensemble Model Based Reinforcement Learning 1 Introduction 2 Related Work 3 Background 4 Method 4.1 Model Discrepancy and Consistency 4.2 Model Learning 4.3 Implementation 5 Experiments 5.1 Comparative Evaluation 5.2 Effects of Consistency Regularization 5.3 Ablation Study 6 Conclusions References Detecting and Learning Against Unknown Opponents for Automated Negotiations 1 Introduction 2 Related Work 3 Preliminaries 3.1 Negotiation Settings 3.2 Bayes Policy Reuse 4 Agent Design 4.1 Deep Reinforcement Learning Based Learning Module 4.2 Policy Reuse Mechanism 5 Experiments 5.1 Experimental Setup 5.2 Performance Against ANAC Winning Agents 5.3 New Opponent Detection and Learning 6 Conclusion References Diversity-Based Trajectory and Goal Selection with Hindsight Experience Replay 1 Introduction 2 Background 2.1 Reinforcement Learning 2.2 Goal-Oriented Reinforcement Learning 2.3 Deep Deterministic Policy Gradient 2.4 Determinantal Point Processes 3 Related Work 4 Methodology 4.1 Diversity-Based Trajectory Selection 4.2 Diversity-Based Goal Selection 5 Experiments 5.1 Environments 5.2 Training Settings 5.3 Benchmark Results 5.4 Ablation Studies 5.5 Time Complexity 6 Conclusion References Off-Policy Training for Truncated TD() Boosted Soft Actor-Critic 1 Introduction 2 Related Work 2.1 TD Learning and Multi-step Methods 2.2 TD() and Eligibility Traces 3 Preliminaries 3.1 MDPs and Temporal Difference Learning 3.2 Multi-step Algorithms and TD() 4 Soft Actor-Critic with Truncated TD () 4.1 Off-Policy Truncated TD() 4.2 Soft Actor-Critic with Truncated TD() 4.3 SAC() Training 5 Experiments 5.1 Evaluation of SAC() 5.2 Ablation Study 6 Discussion References Adaptive Warm-Start MCTS in AlphaZero-Like Deep Reinforcement Learning 1 Introduction 2 Related Work 3 Warm-Start AlphaZero Self-play 3.1 The Algorithm Framework 3.2 MCTS 3.3 MCTS Enhancements 4 Adaptive Warm-Start Switch Method 5 Experimental Setup 6 Results 6.1 MCTS Vs MCTS Enhancements 6.2 Fixed I Tuning 6.3 Adaptive Warm-Start Switch 7 Discussion and Conclusion References Batch-Constraint Inverse Reinforcement Learning 1 Introduction 2 Offline Inverse Reinforcement Learning 3 Method 3.1 Feature Expectation Approximation 3.2 Policy Optimization with BRL 3.3 Batch-Constraint Inverse Reinforcement Learning Algorithm (BCIRL) 4 Experiments 4.1 Standard Control Environments 4.2 Gridworld Example 5 Conclusion References KG-RL: A Knowledge-Guided Reinforcement Learning for Massive Battle Games 1 Introduction 2 Related Work 3 Method 3.1 Rule-Mix 3.2 Plan-Extend 4 Experiment Setup 4.1 Environment 4.2 Human Knowledge Based Module Design 4.3 Experiment Settings 5 Experimental Results 5.1 Battle Game 5.2 Comparison of Training Process 5.3 Model Differences 5.4 The Influence of Different Decisions and Action Modules 5.5 Discussion 6 Conclusion References Vision and Perception A Semi-supervised Defect Detection Method Based on Image Inpainting 1 Introduction 2 Related Work 3 Methodology 3.1 Architecture 3.2 Loss Function 4 Experiments 4.1 Preparations 4.2 Implementation Details 4.3 Results 5 Conclusions References ANF: Attention-Based Noise Filtering Strategy for Unsupervised Few-Shot Classification 1 Introduction 2 Related Work 3 Approach 3.1 Dictionary Noises 3.2 Direct Noise Filter 3.3 Attention-Based Noise Filter 3.4 Dynamic Momentum Updating 4 Experiments 4.1 Datasets 4.2 Implementation Details 4.3 Experimental Results 4.4 Visualization of Filter Results 4.5 Ablation Studies 4.6 Traditional Feature Descriptor 5 Conclusions References Asymmetric Mutual Learning for Unsupervised Cross-Domain Person Re-identification 1 Introduction 2 Related Work 3 Proposed Method 3.1 Structure of Asymmetric Mutual Learning 3.2 Merging Clusters Algorithm 3.3 Similarity Weighted Loss 4 Experiments 4.1 Datasets 4.2 Implementation Details 4.3 Comparison with State-of-the-Art Methods 4.4 Ablation Study 5 Conclusion References Collaborative Positional-Motion Excitation Module for Efficient Action Recognition 1 Introduction 2 Related Work 2.1 Action Recognition 2.2 CNN-Based Approaches 2.3 Temporal Modeling in Action Recognition 2.4 Attention Mechanisms 3 Approach 3.1 Architecture of CPME 3.2 CPME Network 4 Experiments 4.1 Experimental Settings 4.2 Implementation Details 4.3 Improving the Baseline 2D CNN-Approach 4.4 Comparison with the State of the Art 5 Conclusion References Graph Attention Convolutional Network with Motion Tempo Enhancement for Skeleton-Based Action Recognition 1 Introduction 2 Related Work 2.1 GCN for Skeleton Action Recognition 2.2 Motion Tempo Modeling 3 Method 3.1 Multi-neighborhood Graph Attention Module 3.2 Motion Tempo Modeling 4 Experiments 4.1 Datasets 4.2 Training Details 4.3 Ablation Study 4.4 Comparisons with the State-of-the-Art Methods 5 Conclusion References Learning to Synthesize and Remove Rain Unsupervisedly 1 Introduction 2 Related Work 2.1 Single Image Deraining Methods 2.2 Rain Synthesis Methods 2.3 Generative Adversarial Networks 3 SAA-CycleGAN 3.1 Overview 3.2 Deraining Process 3.3 Rain Synthesis Process 3.4 Objective Function 4 Experimental Results 4.1 Implementation Details 4.2 Rain Synthesis Results 4.3 Deraining Results 4.4 Ablation Study 5 Conclusion References Object Bounding Box-Aware Embedding for Point Cloud Instance Segmentation 1 Introduction 2 Related Work 2.1 Deep Learning Methods on Point Cloud 2.2 Instance Segmentation on Point Cloud 3 Method 3.1 Network Framework 3.2 Bounding Box Prediction Branch 3.3 Instance Segmentation Branch 4 Experiments 4.1 Experiment Settings 4.2 Ablation Study 4.3 Comparison with State-of-the-Art Approaches 5 Conclusion References Objects as Extreme Points 1 Introduction 1.1 Key-Point-Based Prediction 1.2 Dense Prediction 1.3 Motivation 2 Related Work 2.1 Anchor-Free Object Detection 2.2 Localization and Classification Spatial Misalignment 2.3 Regression Loss 3 Method 3.1 Positive Sampling with Dynamic Radius 3.2 Network Outputs 3.3 EIoU Loss 3.4 EIoU Predictor 3.5 Optimization 4 Experiments 4.1 Implementation Details 4.2 Ablation Study 4.3 State-of-the-Art Comparisons 5 Conclusion References Occlusion-Aware Facial Expression Recognition Based Region Re-weight Network 1 Introduction 2 Related Work 2.1 FER Methods Against Occlusions 2.2 Sparse Representation 3 Proposed Method 3.1 Overview of Region Re-weight Network 3.2 Occlusion-Aware Module 3.3 Block-Loss Module 4 Experiments 4.1 Datasets 4.2 Implementation Details 4.3 Visualization of the Blocks Selected by OAM 4.4 Ablation Studies Evaluation 4.5 Results and Comparison 5 Conclusion References Online Multi-Object Tracking with Pose-Guided Object Location and Dual Self-Attention Network 1 Introduction 2 Related Work 3 Proposed Method 3.1 Soft-Pose-NMS Object Detection Strategy 3.2 Feature Extraction with Dual Self-Attention Network 3.3 Data Association and Trajectory Management 4 Experiments 4.1 Implementation Details 4.2 Performance on MOT Benchmark Datasets 4.3 Ablation Studies 5 Conclusions References Random Walk Erasing with Attention Calibration for Action Recognition 1 Introduction 2 Related Work 2.1 Video Action Recognition 2.2 Motion Occlusion in Video 3 Approach 3.1 Network Overview 3.2 Random Walk Erasing Module 3.3 Attention Calibration Module 4 Experiments 4.1 Datasets and Implementations 4.2 Main Results 4.3 Ablation Studies 5 Conclusion References RGB-D Based Visual Navigation Using Direction Estimation Module 1 Introduction 2 Related Works 3 Method 3.1 Task Definition 3.2 3D Geometry 3.3 Visual and Spatial Features of Objects 3.4 Direction Estimation Module 3.5 Actor-Critic Policy Network 4 Experiment 4.1 Dataset and Evaluation 4.2 Experiment Setup and Comparison Methods 4.3 Training Details 4.4 Results and Analysis 4.5 Ablation Study 5 Conclusion References Semi-supervised Single Image Deraining with Discrete Wavelet Transform 1 Introduction 2 Related Works 3 Semi-supervised Image Deraining by DWT 3.1 Methodology Overview 3.2 Residual Attentive Network Architecture 3.3 Discriminator by DWT for Semi-supervised Method 4 Experimental Results 4.1 Datasets and Measurements 4.2 Implementation Details 4.3 Results and Analysis 4.4 Ablation Study 5 Conclusion References Simple Light-Weight Network for Human Pose Estimation 1 Introduction 2 Methodology 2.1 Adaptive Convolution 2.2 Light-Weight Block 2.3 Heterogeneous Filters Based Light-Weight Block 2.4 Network Architecture 3 Experiments 3.1 Experiment Setup 3.2 Results 3.3 Ablation Study 3.4 Qualitative Results 4 Conclusion References SIN: Superpixel Interpolation Network 1 Introduction 2 Related Work 2.1 Traditional Superpixel Segmentation 2.2 Superpixel Segmentation Using DNN 2.3 Spatial Connectivity 3 Superpixel Segmentation Method 3.1 Learn Superpixels by Interpolation 3.2 Network Architecture and Loss Function 3.3 Illustration of Spatial Connectivity 4 Experiments 4.1 Comparison with the State-of-the-Arts 4.2 Ablation Study 5 Application 6 Conclusion References SPANet: Spatial and Part-Aware Aggregation Network for 3D Object Detection 1 Introduction 2 Related Work 2.1 Single-Stage Approaches 2.2 Two-Stage Approaches 3 SPANet 3.1 Voxel-Based Backbone 3.2 Spatial Aggregation Network 3.3 Part-Aware Aggregation 3.4 Loss Function 4 Experiments 4.1 Dataset 4.2 Implementation Details 4.3 Results 4.4 Ablation Studies 5 Conclusion References Subspace Enhancement and Colorization Network for Infrared Video Action Recognition 1 Introduction 2 Related Work 2.1 Visible-Based Action Recognition 2.2 Infrared-Based Action Recognition 3 Method 3.1 Framework 3.2 Subspace Enhancement 3.3 Subspace Colorization 3.4 Fusion 4 Experiment and Metrics 4.1 Datasets 4.2 Experimental Settings 4.3 Comparisons with Other Methods 4.4 Ablation Studies 5 Conclusion References Thinking in Patch: Towards Generalizable Forgery Detection with Patch Transformation 1 Introduction 2 Related Work 2.1 Fake Face Generation 2.2 Forgery Detection 3 Approach 3.1 FDPT Architecture for Face Forgery Detection 3.2 Local Subtle Artifacts Learning 3.3 Global Spatial Features Learning 3.4 Fusion Strategy 4 Evaluation 4.1 Experiment Setting 4.2 Ablation Study 4.3 Comparison with Existing Methods 4.4 Generalizability 4.5 Impacts of Image Quality 5 Conclusion References When Distortion Meets Perceptual Quality: A Multi-task Learning Pipeline 1 Introduction 2 Related Work 3 Our MTL-IQA Framework 3.1 Enhanced Edge Fusion Module 3.2 Multi-layer Feature Fusion 3.3 Multi-task Learning for IQA and Classification 3.4 Bayesian Uncertainty-Based Automatically Loss Weighting 4 Experiments 4.1 Experimental Protocol 4.2 Training Strategy 4.3 Evaluation for IQA Task 4.4 Evaluation for Classification Task 4.5 Convergence Evaluation 4.6 Ablation Study 5 Conclusion References Feature Adaption with Predicted Boxes for Oriented Object Detection in Aerial Images 1 Introduction 2 Related Work 2.1 Oriented Object Detection 2.2 Feature Adaption 2.3 Regression Loss 3 Proposed Methods 3.1 Feature Adaption Module 3.2 Decoupled Branch 3.3 Angle Regression Loss 4 Experiments and Analysis 4.1 Datasets 4.2 Implementation Details 4.3 Ablation Studies 4.4 Visualization on DOTA 5 Conclusion References Few-Shot Crowd Counting via Self-supervised Learning 1 Introduction 2 Related Work 3 Our Method 3.1 Crowd Count Ranking 3.2 Crowd Distance Ranking 3.3 Model 3.4 Training 4 Experiments 4.1 Dataset 4.2 Settings 4.3 Comparing Methods 4.4 Results 4.5 Feature Visualization 5 Conclusion References Low-Rank Orthonormal Analysis Dictionary Learning for Image Classification 1 Introduction 2 Related Work 2.1 Dictionary Learning 2.2 Low-Rank Representation 3 The LR-OADL Model 4 Optimization 5 Experiments 5.1 Results and Analysis 6 Conclusions References MRAC-Net: Multi-resolution Anisotropic Convolutional Network for 3D Point Cloud Completion 1 Introduction 2 Related Work 2.1 3D Shape Completion 2.2 Convolution-Based Networks 3 Approach 3.1 Anisotropic Convolutional Encoder 3.2 Combined Pyramid Decoder 3.3 Loss Function 4 Experiments 4.1 Data Generation 4.2 Implementation Detail 4.3 Results 4.4 Ablation Study 5 Conclusion References Nonlinear Parametric Transformation and Generation of Images Based on a Network with the CWNL Layer 1 Introduction 2 Mathematical Model of the CWNL Layer with External Control 2.1 Training Methodology of the CWNL Layer 3 Experiments 3.1 Nonlinear Transformation 3.2 Nonlinear Generation 4 Conclusions References PupilFace: A Cascaded Face Detection and Location Network Fusing Attention 1 Introduction 2 Related Work 3 PupilFace 3.1 Nonlinearities and Multi-task Loss 3.2 Design of Fusing Attention Networks 4 Experiments 4.1 Dataset 4.2 Implementation Details 4.3 Experimental Results 5 Conclusion References Author Index