Image Analysis and Processing – ICIAP 2022: 21st International Conference, Lecce, Italy, May 23–27, 2022, Proceedings, Part III

Preface
Organization
Contents – Part III
Pattern Recognition and Machine Learning
Hangul Fonts Dataset: A Hierarchical and Compositional Dataset for Investigating Learned Representations
	1 Introduction
	2 The Hangul Fonts Dataset
	3 Methods
	4 Results
	5 Discussion
	References
Out-of-Distribution Detection Using Outlier Detection Methods
	1 Introduction
	2 Related Literature
		2.1 Out-of-Distribution Detection
		2.2 Outlier Detection
	3 Method
	4 Experiment Setup
	5 Results
	6 Discussion and Conclusion
	References
Relaxation Labeling Meets GANs: Solving Jigsaw Puzzles with Missing Borders
	1 Introduction
	2 Related Works
	3 Model
		3.1 Border Extension
		3.2 Pairwise Compatibility
		3.3 Relaxation Labeling Puzzle Solver
	4 Experiments and Results
	5 Conclusion
	References
Computationally Efficient Rehearsal for Online Continual Learning
	1 Introduction
	2 Related Work
	3 Online Continual Learning
		3.1 Scenario
		3.2 The Proposed Online Rehearsal Method Framework
	4 Alternative Rehearsal Strategies
		4.1 Continuous Rehearsal
		4.2 Drift Activated Rehearsal
		4.3 Dynamic Number of Training Iterations
		4.4 Iterate Until Convergence
		4.5 Adjust Learning Rate
	5 Experiments
		5.1 Experimental Setup
		5.2 Experiment 1: The Need for Continuous Rehearsal
		5.3 Experiment 2: Choosing the Best Rehearsal Strategy
	6 Conclusions
	References
Recurrent Vision Transformer for Solving Visual Reasoning Problems
	1 Introduction
	2 Related Work
	3 The Recurrent Vision Transformer Model
	4 Experiments
		4.1 Dataset
		4.2 Setup
		4.3 Results
		4.4 Ablation Study
	5 Conclusions
	References
Metric Learning-Based Unsupervised Domain Adaptation for 3D Skeleton Hand Activities Categorization
	1 Introduction
	2 Related Work
		2.1 Unsupervised Domain Adaptation (UDA)
		2.2 Metric Learning
	3 Proposed Method
		3.1 Supervised Learning for Discriminative Features Space Creation
		3.2 Unlabeled Activities Clustering
		3.3 Consensus Clustering
		3.4 Automated Penalty-Based Metric Learning (APML) Loss Function
	4 Experiments
		4.1 Data Set
		4.2 Evaluation Protocols
		4.3 Implementation Details
		4.4 State-of-the-Art Comparison
		4.5 Unlabeled Activities Clustering Analysis
	5 Conclusion
	References
Using Random Forest Distances for Outlier Detection
	1 Introduction
	2 Methodology
		2.1 Step 1: Building an ERF
		2.2 Step 2: Extracting the Distance Matrix D
		2.3 Step 3: Distance-Based Outlier Detection
	3 Experimental Evaluation
		3.1 Experimental Details
		3.2 ERF Parametrization
		3.3 Comparison of Outlier Detectors
		3.4 Comparison of Distance Measures
	4 Conclusions
	References
Case Study on the Use of the SafeML Approach in Training Autonomous Driving Vehicles
	1 Introduction
	2 Robustness and Safety of Machine Learning Algorithms in Autonomous Driving
	3 SafeML Approach
	4 NuImage Data Set
	5 Experimental Setup
		5.1 Data Set Generation
		5.2 Alexnet with Spartial Pyramid Pooling Layer
		5.3 Distance Measurements
	6 Results
	7 Discussion
	8 Conclusion
	References
User-Biased Food Recognition for Health Monitoring
	1 Introduction
	2 Related Work
	3 Proposed Method
		3.1 Data Acquisition and Annotation
		3.2 Proposed FoodRec Architecture
		3.3 User Data Annotation
		3.4 Data Augmentation
	4 Experimental Results
		4.1 Comparative Evaluation
	5 Conclusion
	References
Multi-view Spectral Clustering via Integrating Label and Data Graph Learning
	1 Introduction
	2 Preliminaries and Related Work
		2.1 Notations
		2.2 Related Work
		2.3 Review of the (NESE) Method
	3 Proposed Approach
		3.1 Optimization
	4 Performance Evaluation
		4.1 Experimental Setup
		4.2 Experimental Results
		4.3 Parameter Sensitivity
		4.4 Analysis of Results
		4.5 Convergence Study
	5 Conclusion
	References
Distance-Based Random Forest Clustering with Missing Data
	1 Introduction
	2 Random Forest Clustering with RatioRF
		2.1 The RatioRF Distance
		2.2 The Complete Random Forest Clustering Procedure
	3 Dealing with Missing Data
		3.1 Computing RatioRF with Missing Data
		3.2 Training Trees with Missing Data
	4 Experimental Evaluation
		4.1 Experimental Details
		4.2 Results and Discussion
	5 Conclusions
	References
Video Analysis and Understanding
Unsupervised Person Re-identification Based on Skeleton Joints Using Graph Convolutional Networks
	1 Introduction
	2 Related Work
	3 Proposed Method
		3.1 Preliminary
		3.2 Joint Graph Construction
		3.3 GCN Module
		3.4 Temporal Attention Module
		3.5 Clustering Module
		3.6 Loss Functions
	4 Experimental Results
		4.1 Datasets
		4.2 Experimental Settings
		4.3 Comparison with the State-of-the-Art Methods
		4.4 Algorithm Analysis
	5 Conclusion
	References
Keyframe Insights into Real-Time Video Tagging of Compressed UHD Content
	1 Introduction
	2 Related Work
	3 The Impact of Increasing Video Resolutions on Video Analysis Tasks
		3.1 Coding Structure of a Video
		3.2 Keyframes to the Rescue
	4 Keyframe-Based Video Genre Tagging
		4.1 Training Video Genre Tagging
		4.2 Predicting Video Genres from Keyframes at Inference Time
	5 Experimental Validation
		5.1 Frame Extraction Speed Comparison
		5.2 Frame Genre Tagging Speed
		5.3 Quantitative Evaluation of the Proposed Genre Tagging
		5.4 Qualitative Results
	6 Conclusions and Future Work
	References
Exploring the Use of Efficient Projection Kernels for Motion Saliency Estimation
	1 Introduction
	2 Gray-Code Filter Kernels
		2.1 Efficient Filtering with GCK
		2.2 3D Gray-Code Kernels
	3 Proposed Method for Motion Saliency Estimation
		3.1 Projection Module: Filtering and Pooling
		3.2 GCKs Attention Module
	4 Experimental Results
		4.1 Datasets and Evaluation Metrics
		4.2 Method Assessment
		4.3 Evaluations on VOS Datasets
		4.4 Computational Analysis
	5 Discussion
	References
FirstPiano: A New Egocentric Hand Action Dataset Oriented Towards Augmented Reality Applications
	1 Introduction
	2 Related Work
	3 FirstPiano: Egocentric Dataset of Piano Interaction
		3.1 Dataset Overview
		3.2 Sensors and Acquisition Modalities
	4 Benchmark Evaluation
		4.1 Method: 2D Deep Video Capsule Network
		4.2 Experiments
	5 Conclusion
	References
Learning Video Retrieval Models with Relevance-Aware Online Mining
	1 Introduction
	2 Related Work
	3 Training a Video Retrieval Model with Contrastive Loss and Mining
		3.1 Online Hard Negative Mining
	4 Proposed Method: Relevance-Aware Online Mining
		4.1 Relevance
		4.2 Relevance-Aware Online Hard Negative Mining
		4.3 Relevance-Aware Online Hard Positive Mining
	5 Results
		5.1 Analysis of Hard Negatives: Relevance Distribution
		5.2 Influence of the Threshold  on the Proposed Techniques
		5.3 Comparison with State-of-the-Art
	6 Conclusions
	References
Foreground Detection Using an Attention Module and a Video Encoding
	1 Introduction
	2 Related Work
	3 Methodology
	4 Evaluation
		4.1 CDnet2014 Dataset
		4.2 Evaluation Metrics
		4.3 Trainable Parameters
		4.4 Results
		4.5 Discussion
	5 Conclusions
	References
Test-Time Adaptation for Egocentric Action Recognition
	1 Introduction
	2 Related Works
		2.1 Egocentric Action Recognition
		2.2 Cross-domain Action Recognition
	3 Problem Formulation
	4 Test-Time Adaptation for Action Recognition
		4.1 Multi-modal Test-Time Adaptation
		4.2 Class Relative Losses
	5 Experimental Results
		5.1 Experimental Setting
		5.2 Results
	6 Conclusions
	References
Combining EfficientNet and Vision Transformers for Video Deepfake Detection
	1 Introduction
	2 Related Works
		2.1 Deepfake Generation
		2.2 Deepfake Detection
	3 Method
	4 Experiments
		4.1 Datasets and Face Extraction
		4.2 Training
		4.3 Inference
		4.4 Results
	5 Conclusions
	References
Human Action Recognition with Transformers
	1 Introduction
	2 Related Works
	3 Proposed Methodology
		3.1 Channel Separated Convolutional Network
	4 Datasets
		4.1 HMDB-51
		4.2 UCF-101
		4.3 IG-65M
	5 Experimental Results
		5.1 Ablation Study
		5.2 Performance Comparison
	6 Conclusions
	References
Decontextualized I3D ConvNet for Ultra-Distance Runners Performance Analysis at a Glance
	1 Introduction
	2 Related Work
	3 Runner Performance Pipeline
		3.1 Runners Tracking and Segmentation
		3.2 Runners Features Extraction
	4 Experimental Evaluation
		4.1 Experimental Setup
		4.2 Experimental Results
	5 Conclusions
	References
Densification of Sparse Optical Flow Using Edges Information
	1 Introduction
	2 The Proposed Approach
	3 Experimental Results
		3.1 Test 1: Objective Impact of Densification
		3.2 Test 2: Benefits of Densification in Clustering
		3.3 Test 3: Densification Impact on Odometry
	4 Conclusion
	References
Cycle Consistency Based Method for Learning Disentangled Representation for Stochastic Video Prediction
	1 Introduction
	2 Related Work
	3 Our Approach
		3.1 VAE-GAN Models
		3.2 Cycle Consistency Loss
		3.3 Architecture Details
	4 Experiments
		4.1 Stochastic Moving MNIST
		4.2 BAIR Robot Pushing Dataset
	5 Conclusion
	References
SeeFar: Vehicle Speed Estimation and Flow Analysis from a Moving UAV
	1 Introduction
	2 Related Work
		2.1 Object Detection
		2.2 Multi-Object Tracking
		2.3 Visual Speed Estimation
	3 Methods
		3.1 System Architecture
		3.2 Detector and Tracker
		3.3 Speed Estimator
		3.4 Flow Analysor
	4 Experiments and Analysis
		4.1 Speed Estimator
		4.2 Flow Analysor
	5 Conclusion
	References
Spatial-Temporal Autoencoder with Attention Network for Video Compression
	1 Introduction
	2 Related Work
		2.1 Learned Image Compression
		2.2 Learned Video Compression
		2.3 Attention Mechanism
	3 The Proposed STVC Approach
		3.1 Framework
		3.2 Spatial-Temporal Priors with an Attention Module (STPA)
		3.3 Loss Function
	4 Experiments
		4.1 Training the Proposed Network
		4.2 Evaluating the Performance
	5 Conclusion and Future Work
	References
On the Evaluation of Video-Based Crowd Counting Models
	1 Introduction
	2 Background
	3 A Categorisation of Video-Based Crowd Counting Models
	4 Evaluation of the Temporal Contribution
	5 Experiments
		5.1 Experimental Setting
		5.2 Results
	6 Conclusions and Future Work
	References
Frame-Wise Action Recognition Training Framework for Skeleton-Based Anomaly Behavior Detection
	1 Introduction
	2 Related Work
		2.1 Skeleton-Based Action Recognition Models
		2.2 Skeleton-Based Anomaly Behavior Detection
	3 Method
		3.1 Selecting Frames for Frame-Wise Action Recognition Training
		3.2 Frame-Wise AGCN
		3.3 Anomaly Behavior Detection Method
	4 Experiments
		4.1 Datasets and Evaluation
		4.2 Experimental Setting
		4.3 Results
		4.4 Ablation Study
	5 Conclusion
	References
The Automated Temporal Analysis of Gaze Following in a Visual Tracking Task
	1 Introduction
	2 Extraction of Data and Pre-processing
		2.1 EYEDIAP Video Dataset
		2.2 Extraction of Gaze and Object Displacement Data
		2.3 Pre-processing
	3 Time-Series Analysis Between the Gaze and Object Displacement
		3.1 Statistical Correlation
		3.2 Dynamic Time Warping (DTW)
	4 Results and Discussion
		4.1 Multivariate Time-Series Analysis Metrics
		4.2 Intragroup Comparison of Temporal Characteristics
		4.3 Intragroup Comparison of Response to Horizontal and Vertical Motion
	5 Conclusion and Future Works
	References
Untrimmed Action Anticipation
	1 Introduction
	2 Related Works
	3 Problem Formulation
	4 Evaluation Protocol
	5 Methods
		5.1 Controlled Experiments
		5.2 Baselines Based on Trimmed Action Anticipation
	6 Experimental Settings and Results
		6.1 Results of the Controlled Experiments
		6.2 Trimmed Action Anticipation baseline Results
	7 Conclusion
	References
Forecasting Future Instance Segmentation with Learned Optical Flow and Warping
	1 Introduction
	2 Related Work
	3 Our Approach
		3.1 Optical Flow Forecasting
		3.2 Instance Mask Forecasting
	4 Experiments
		4.1 Implementation Details and Ablation Study
	5 Results
	6 Conclusions
	References
Depth-Aware Multi-object Tracking in Spherical Videos
	1 Introduction
	2 Related Work
	3 Modeling Targets' Locations on Spherical Images
		3.1 360 Videos and Equirectangular Images
		3.2 Estimating Targets' Distances and Locations
	4 Proposed MOT Algorithm
		4.1 Pedestrian Detection
		4.2 Modeling and Predicting Targets' Locations
		4.3 Data Association
		4.4 Trackers' Birth, Death and Updating
	5 Experimental Results
		5.1 Results
	6 Conclusions and Future Work
	References
FasterVideo: Efficient Online Joint Object Detection and Tracking
	1 Introduction
	2 Related Work
		2.1 Video Object Detection
		2.2 Multi-object Tracking
		2.3 Joint Object Detection and Tracking
	3 Proposed Method
		3.1 Video Object Detection
		3.2 Embeddings Learning
		3.3 Data Association and Tracking
	4 Experimental Analysis
		4.1 KITTI Benchmark
		4.2 MOT Benchmark
		4.3 Inference Time Analysis
	5 Conclusion and Future Work
	References
A Large-scale TV Dataset for Partial Video Copy Detection
	1 Introduction
	2 Related Work
	3 STVD: A Large-scale TV Dataset
		3.1 TV Video Capture (C1)
		3.2 Video Detection (C2)
		3.3 Video Degradation (C3)
	4 Experiments
		4.1 Dataset and Groundtruthing
		4.2 Performance Evaluation
	5 Conclusions and Perspectives
	References
Poker Bluff Detection Dataset Based on Facial Analysis
	1 Introduction
	2 Related Works
	3 Methods/Technical Details
		3.1 Parsing Game Information
		3.2 Generating Bluff Labels
		3.3 Dataset Details
	4 Bluff Detection Analysis
	5 Conclusion and Future Work
	References
Engagement Detection with Multi-Task Training in E-Learning Environments
	1 Introduction
	2 Related Works
	3 Architectural Design for Engagement Detection with Multi-Task Training
		3.1 Feature Extraction
		3.2 Feature Aggregation over Time Windows
		3.3 Sequence Modeling Combined with Multi-Tasking
	4 Experimental Results
		4.1 Dataset
		4.2 Experimental Setup and Hyperparameter Settings
		4.3 Performance Evaluation
		4.4 Qualitative Results
	5 Conclusion
	References
Special Session
Ship Detection and Tracking Based on a Custom Aerial Dataset
	1 Introduction
		1.1 MARIN Project
		1.2 Contributions
		1.3 Outline of the Paper
	2 Background
		2.1 You Only Look Once (YOLO)
		2.2 Deep Simple Online and Real Time Tracking with Convolutional Neural Networks
	3 Dataset Construction
		3.1 Dataset Sources
		3.2 Data Preparation
		3.3 Data Splitting
	4 Experimental Results and Analysis
		4.1 Experimental Environment and Configuration
		4.2 Training Results
		4.3 Detection and Tracking Results
	5 Conclusion
	References
Prediction of Fish Location by Combining Fisheries Data and Sea Bottom Temperature Forecasting
	1 Introduction
	2 Related Works
	3 Proposed Framework
		3.1 Datasets Used
		3.2 Predicting Bottom Temperature with Deep Learning
		3.3 Fish Prediction with Gradient Boosting
	4 Experimental Settings
		4.1 Train-Test-Evaluation Splits
		4.2 Hyperparameters
		4.3 Metrics
	5 Performance Evaluation
		5.1 Temperature Forecasting Performance
		5.2 Influence of Bottom Temperature for Fish Prediction
		5.3 Performance of the Complete Framework
	6 Discussion and Conclusion
	References
Robust Human-Identifiable Markers for Absolute Relocalization of Underwater Robots in Marine Data Science Applications
	1 Introduction
	2 Related Work
	3 Robust Marker Design
	4 Experiments
	5 Conclusion
	References
Towards Cross Domain Transfer Learning for Underwater Correspondence Search
	1 Introduction
	2 D2Net Revisited for Underwater Training
		2.1 Triplet Margin Ranking Loss with Underwater Images
		2.2 Robustness Against Basic Image Transformations
		2.3 Consequences for Underwater Matching
	3 Cross-domain Transfer Learning
		3.1 Physical Simulation of Underwater Effects
		3.2 Model Fine-tuning
	4 Evaluation
	5 Conclusion
	References
On the Evaluation of Two Methods Applied to the Morphometry of Linear Dunes
	1 Introduction
	2 Materials and Methods
	3 Results and Discussion
	4 Conclusion
	References
Underwater Image Enhancement Using Pre-trained Transformer
	1 Introduction
	2 Related Work
	3 Methodology
	4 Experiment
	5 Results
	6 Conclusions
	References
Author Index