Frontiers of Computer Vision: 28th International Workshop, IW-FCV 2022, Hiroshima, Japan, February 21–22, 2022, Revised Selected Papers (Communications in Computer and Information Science)

This book constitutes refereed proceedings of the 28th International Workshop on Frontiers of Computer Vision, IW-FCV 2022, held in Hiroshima, Japan, in February 2022. Due to the COVID-19 pandemic the conference was held online. The 24 full papers presented in this volume were thoroughly reviewed and selected from 63 submissions. The papers are organized according to the following topics: ​camera, 3D, and imaging; learning algorithm; object detection/segmentation; recognition/generation. Preface Organization Contents Camera, 3D, and Imaging 6D Pose Estimation of Transparent Objects Using Synthetic Data 1 Introduction 2 Related Works 2.1 Transparent Object Recognition and Pose Estimation 2.2 Synthetic Data for Training 2.3 Object 6D Pose Estimation 3 Proposed Pose Estimation of Transparent Objects 3.1 Data Generation 3.2 Network Architecture 3.3 Detection and Pose Estimation 4 Experimental Results 4.1 Setting and Dataset 4.2 Evaluation Metrics 4.3 Results of Synthetic Image 4.4 Results of Real Image 4.5 Discussion and Limitation 5 Conclusion References Color Exaggeration for Dichromats Using Weighted Edge 1 Introduction 2 Hue for Dichromats 3 Color Enhancement for Dichromat 4 Experiment 5 Conclusion References Uncalibrated Photometric Stereo Using Superquadrics with Texture Estimation 1 Introduction 2 Formulation and Algorithms 2.1 Minimization on Two Sets of Parameters 2.2 Three-Step Algorithm 2.3 Jacobian 3 Image Formation Model 3.1 Superquadric Surfaces and Coordinate Systems 3.2 Rendering Superquadric Surfaces 3.3 Shading Model 3.4 Two-Stage Shading and Jacobian 3.5 Dealing with Contour Mismatch 3.6 Contour of a Superquadric Surface and a Cast Shadow 4 Experimental Results 4.1 Datasets and Implementation 4.2 Entries of 4.3 Initial Reflectance Map 4.4 Simulation 4.5 Grapes 5 Conclusions References 3D Shape Reconstruction of Japanese Traditional Puppet Head from CT Images 1 Introduction 2 3D Shape Reconstruction from CT Images 2.1 Extracted Regions in CT Images 2.2 Extracting Materials Based on Histogram 2.3 Precise Extraction by Graph Cut Method 2.4 Extraction of Inner Parts 2.5 Extraction by Machine Learning 3 Experiments 3.1 Puppet Head and CT Images 3.2 Experimental Results of Puppet Shape Extraction 3.3 Experimental Results of Inner Parts Extraction 3.4 Experimental Results of Puppet Shape Extraction by U-Net 4 Conclusion References Multi-band Photometric Stereo Using Random Sampling of Channels and Pixels 1 Introduction 2 Multispectral Color Photometric Stereo Method 2.1 Image Formulation 2.2 Multiple Albedo with 4 Channels and 8 Pixels 2.3 Computing Albedo and Surface Normal 3 Experiment 4 Conclusion References Online Illumination Planning for Shadow-Robust Photometric Stereo 1 Introduction 2 Related Work 3 Proposed Method 3.1 Overview 3.2 Estimating Surface Normals from Images with Shadows 3.3 Finding Pixel with Worst Accuracy 3.4 Finding Optimal Light Source Direction 4 Experiments 4.1 Synthetic Images 4.2 Real Images 5 Conclusion and Future Work References Learning Algorithm Decomposition of Invariant and Variant Features by Using Convolutional Autoencoder 1 Introduction 2 Related Works 2.1 Autoencoder 2.2 Invariant Features 3 Proposed Network Architecture 3.1 Combination of Classifier and Autoencoder 3.2 Combination of Two Autoencoders 4 Experiments 4.1 Combination of Classifier and Autoencoder 4.2 Combination of Two Autoencoders 4.3 Image Reconstruction 5 Conclusion and Future Works References Deep Automatic Control of Learning Rates for GANs 1 Introduction 2 Related Work 3 Prediction of Training Success Degree 3.1 Training Success Degree of Conditional GAN 3.2 Training Success Degree of Unconditional GAN 3.3 Prediction of Training Success Degree 3.4 Training of the Training Success Predictor 4 Learning Rate Control 4.1 Estimating the Optimal Learning Rate 4.2 Learning Rate Control in GAN Training 5 Experiments 5.1 Prediction of Training Success Degree in Conditional GAN 5.2 Prediction of Training Success Degree in Unconditional GAN 5.3 Learning Rate Control 6 Conclusion References Multimodal Pseudo-Labeling Under Various Shooting Conditions: Case Study on RGB and IR Images 1 Introduction 2 Related Work 2.1 Object Detection 2.2 Multimodal Dataset 2.3 Pseudo-Labeling and Self-training 2.4 Problem of Existing Methods 3 Method 3.1 Prerequisite 3.2 Overall Procedure 3.3 Transformation Between Two Modalities 4 Experiments 4.1 Dataset 4.2 Implementation Details 4.3 Quantitative Results 4.4 Qualitative Results 5 Conclusion References Object Detection/Segmentation Convolutional Neural Network Design for Eye Detection Under Low-Illumination 1 Introduction 2 Related Work 3 Methodology 3.1 Feature Extraction Module 3.2 Detection Module 3.3 Loss Function 4 Experiments 4.1 Dataset Preparation 4.2 Experimental Setup 4.3 Experimental Result 4.4 Ablation Study 5 Conclusion References Pedestrian Head Detection and Tracking via Global Vision Transformer 1 Introduction 2 Related Works 3 Methodology 4 Experiments 4.1 Dataset and Evaluation Metrics 4.2 Implementation Details 5 Results 6 Conclusion References Proposal of a Method to Identify Vascular Endothelial Cells from Images of Mouse Myocardial Tissue 1 Introduction 2 Related Work 3 Method 3.1 Regional Extraction of Cell Nuclei and Vascular Endothelium 3.2 Determining the Surface of the Cell Nucleus 3.3 Determining Whether the Nuclei is Covered with Vascular Endothelium 3.4 Classification of the Cell Nuclei Using the Percentage of Coverage 4 Experiment 4.1 Data Sets Used 4.2 Experimental Details 5 Results 6 Conclusions and Future Work References Improved Facial Keypoint Regression Using Attention Modules 1 Introduction 2 Proposed Method 2.1 Keypoint Regression 2.2 Proposed Model 2.3 Ground Truth Masks Encoding and Loss Functions 2.4 Attention Modules 3 Experiments 3.1 Training 3.2 Qualitative Results 3.3 Quantitative Results 4 Conclusion References Video Object Segmentation Based on Guided Feature Transfer Learning 1 Introduction 2 Related Work 3 Proposed Method 3.1 Guided Feature Modulation Module (GFM) 3.2 Generative Appearance Module (GAM) 3.3 Guided Pooled Decoder 3.4 Video Object Segmentation Architecture 3.5 Network Training Details 4 Experiments and Evaluations 4.1 Datasets 4.2 Comparison Methods 4.3 Qualitative Analysis 4.4 Ablation Study 5 Conclusion References Improvement of On-Road Object Detection Using Inter-region and Intra-region Attention for Faster R-CNN 1 Introduction 2 Related Methods 2.1 Faster R-CNN 2.2 Region Proposal Network 2.3 Attention Module 2.4 Lambda Layer 3 Proposed Methods 3.1 Self Attention for Final Features Between High-Confidence Regions 3.2 Attention for Final Features Between Medium- and High-Confidence Regions 3.3 Lambda Layer for ROI Feature Maps 4 Experiments 4.1 Experimental Setup 4.2 Comparison Between Faster R-CNN and Proposed Methods 4.3 Comparison Between SA-FRCNN and STA-FRCNN 4.4 Comparison Between Lambda-FRCNN and Other Methods 4.5 Comparison in Average Recall 5 Conclusions References Deep Segmentation Network Without Mask Image Supervision for 2D Image Registration 1 Introduction 2 Our Deep Segmentation Network with 2D Image Registration 2.1 Overview 2.2 Training Image Pairs 2.3 Deep Segmentation Network Module 2.4 Inference 3 Experiments 3.1 Dataset 3.2 Implementation and Performance Metrics 3.3 Comparison with SuperPoint ch16DeTonesps2018 + SuperGlue ch16Sarlinsps2020 3.4 Comparison with Existing Segmentation Networks 3.5 Performance Without Features Extracted in 2D Image Registration Module 3.6 Comparison with the Other 2D Image Registration Technique 4 Conclusions References Lightweight Encoder-Decoder Architecture for Foot Ulcer Segmentation 1 Introduction 2 Related Work 2.1 Classical Segmentation Methods 2.2 Deep Learning-Based Segmentation Methods 2.3 Attention Mechanisms 3 Proposed Method 3.1 Model Overview 3.2 Loss Function 3.3 Residual Attention Block 3.4 Experimental Setup 4 Experiments 4.1 Dataset 4.2 Evaluation Metrics 4.3 Comparison with Baseline Model 4.4 Comparison with Challenge Records 5 Conclusion References Recognition/Generation Implementation of Digital Transformation for Korean Traditional Heritage 1 Introduction 2 Attaining High-Quality Digital Heritage 2.1 Enhancement of Existing Data 2.2 Attaining High Quality 2D/3D Traditional Heritage Assets 3 Building an Intelligent Curation Platform 3.1 Tagging Information and Analyzation 3.2 A Web-based Intuitive Platform 3.3 Demonstration and Application 4 Conclusion References Facial Mask Region Completion Using StyleGAN2 with a Substitute Face of the Same Person 1 Introduction 2 Proposed Method 2.1 Overview 2.2 Alignment of Face Images 2.3 Processing by StyleGAN2 2.4 Completion with Poisson Image Editing 3 Experiments 3.1 Experiment 1 3.2 Experiment 2 4 Conclusion References Generation of Omnidirectional Image Without Photographer 1 Introduction 2 Proposed Method 2.1 Overview 2.2 Calculation of Translation 2.3 Unifying the Appearance of Multiple Images 2.4 Image Composition by Image Selection 3 Experiments 3.1 Scene A: Indoor Scene 3.2 Scene B: Outdoor Scene 3.3 Scene C: Narrow Indoor Scene 3.4 Discussion 4 Conclusion References Optimization of Re-ranking Based on k-Reciprocal for Vehicle Re-identification 1 Introduction 2 Related Works 2.1 Vehicle Re-identification 2.2 Re-ranking 3 Proposed Method 3.1 Baseline Model 3.2 k-reciprocal Encoding 3.3 Local Query Expansion with Weight 4 Experiment 4.1 Evaluation Metric 4.2 Data Sets and Implementation Details 4.3 Experiments on VeRi-776 and VehicleID 5 Conclusion References Sequence Recognition of Indoor Tennis Actions Using Transfer Learning and Long Short-Term Memory 1 Introduction 2 Literature Review 3 Workflow and Proposed Architecture 4 Results and Observation 4.1 System Configuration 4.2 Performance Analysis of Combined Xception–LSTM Model 4.3 Results of Pre-trained CNN Features and LSTM 4.4 Results of the Best Model 5 Conclusion References Multi-region Based Radial GCN Algorithm for Human Action Recognition 1 Introduction 2 Related Studies 3 MRGCN Algorithm 3.1 Input Data Generation Procedure 3.2 Design of Graph Structure 4 Experiment 4.1 Experimental Environment and Structure of Neural Network 4.2 Experimental Results 4.3 The Experiment of Real-Time Action Recognition System 5 Conclusion References Impression Estimation Model of 3D Objects Using Multi-View Convolutional Neural Network 1 Introduction 2 Previous Research 3 Modeling the Relationships Between Visual Impressions and Physical Characteristics 4 Building an Impression Estimation Model 4.1 Data Set 4.2 Training 5 Results and Discussions 5.1 Overall Performance 5.2 Comparison of the Proposed Method with Each Comparison Method 5.3 Comparison of Impression Estimates for Each Object Category 5.4 Relationship Between Estimated Evaluation Value and 3D Models 6 Conclusion References Author Index