Pattern Recognition: 5th Asian Conference, ACPR 2019, Auckland, New Zealand, November 26–29, 2019, Revised Selected Papers, Part I (Lecture Notes in Computer Science, 12046)

Preface
Organization
Contents – Part I
Contents – Part II
Classification
Integrating Domain Knowledge: Using Hierarchies to Improve Deep Classifiers
1 Introduction
2 Related Work
3 Method
3.1 Class Hierarchy
3.2 Probabilistic Model
3.3 Inference
3.4 Training
4 Experiments
4.1 Experimental Setup
4.2 Overall Improvement—ImageNet
4.3 Speedup—CIFAR-100
4.4 Fine-Grained Recognition—NABirds
4.5 Overview and Discussion
5 Conclusion
References
Label-Smooth Learning for Fine-Grained Visual Categorization
1 Introduction
2 Related Work
2.1 Fine-Grained Visual Categorization
2.2 Label-Smooth Leaning
3 Approach
3.1 Label-Smoothing Learning
3.2 Cross-Category Cross-Semantic Regularization
4 Experiments
4.1 Fine-Grained Visual Categorization (FGVC) Datasets
4.2 Implementation
4.3 Ablation Studies
4.4 Comparison with State-of-the-Art
5 Conclusion
References
ForestNet – Automatic Design of Sparse Multilayer Perceptron Network Architectures Using Ensembles of Randomized Trees
1 Introduction
2 Related Work
3 Model to Build a Sparse Multilayer Pereceptron Network
3.1 Building the Ensemble of Randomized Trees
3.2 Defining the SMLP Network Architecture
3.3 Initialization of the Weights
4 Experiment Settings
4.1 Hyperparameter Settings for Tree Ensembles
4.2 Hyperparameter Settings for Networks
5 Results and Discussion
5.1 Classification Accuracy
5.2 Amount of Connections
5.3 Amount of Hidden Neurons
5.4 Training and Validation Losses
5.5 Discussion
6 Conclusion and Future Work
References
Clustering-Based Adaptive Dropout for CNN-Based Classification
1 Introduction
2 The Proposed Algorithm
2.1 Clustering-Based Dropout with Adaptive Probability
2.2 Clustering Algorithm and Network Configuration
2.3 Implementation Details
3 Experimental Results
4 Conclusion
References
Action and Video and Motion
Real-Time Detection and Tracking Using Hybrid DNNs and Space-Aware Color Feature: From Algorithm to System
1 Introduction
2 Related Work
2.1 Multiple Object Tracking
2.2 Color Feature
3 Whole Framework
4 Detection
5 Matching and Tracking
5.1 Similarity Sorting-Based Matching
5.2 Space-Aware Color-Based Bounding Box Refinement
5.3 Retire-Recall Mechanism
5.4 Whole Matching and Tracking Algorithm
6 Experiment Evaluation
6.1 Overall Performance
6.2 Technique Effect
6.3 Platform-Specific Exploration
6.4 Discussion
7 Conclusion
References
Continuous Motion Numeral Recognition Using RNN Architecture in Air-Writing Environment
1 Introduction
2 Related Works
3 Proposed Methodology
3.1 Motion Tracking Method
3.2 Numeral Recognition
4 Results and Discussion
4.1 Data Collection
4.2 Experimental Setup
4.3 Result
4.4 Error Analysis
5 Conclusion
References
Using Motion Compensation and Matrix Completion Algorithm to Remove Rain Streaks and Snow for Video Sequences
1 Introduction
2 Challenges
3 The Proposed Method
3.1 Rain Streaks Detection
3.2 Rain Map Refinement
3.3 Rain Streaks Reconstruction
4 Experimental Results
4.1 Synthetic Video Sequences
4.2 Real Video Sequences
4.3 Challenge Tasks
5 Conclusion
References
Assessing the Impact of Video Compression on Background Subtraction
1 Introduction
2 Related Work
3 Problem Formulation
4 Study Design
4.1 Background Subtraction
4.2 Dataset
4.3 Quality Metrics for Videos
4.4 Performance Metric for Background Subtraction Algorithms
5 Results
5.1 Relation Between Percentage Drop in Quality and Performance
5.2 Identify the Encoding Parameters to Employ in Surveillance Scenarios
5.3 Predicting Performance Based on Video Quality
6 Conclusion
References
Object Detection and Anomaly Detection
Learning Motion Regularity for Temporal Video Segmentation and Anomaly Detection
1 Introduction
2 Related Work
3 Anomaly Detection Method
3.1 Temporal Video Segmentation
3.2 Anomaly Detection
4 Experiments
4.1 Dataset
4.2 Implementation Details
4.3 Results
4.4 Analysis of the Impact of Temporal Video Segmentation
5 Conclusion
References
A New Forged Handwriting Detection Method Based on Fourier Spectral Density and Variation
1 Introduction
2 Related Work
3 Proposed Method
3.1 Fourier Spectral Distributions
3.2 Spectral Density and Variance Based Features for Forged Handwriting Classification
4 Experimental Results
4.1 Experiments on Forged Handwriting Word Detection
4.2 Experiments on Forged Caption Text Detection in Video Images
4.3 Experiments on Forged IMEI Number Detection
5 Conclusion and Future Work
References
Robust Pedestrian Detection: Faster Deployments with Fusion of Models
1 Introduction
2 Related Works
3 Methods
3.1 State-of-the-Art Object Detectors
3.2 Non-maxima Suppression
3.3 Detectors Divergences
3.4 Missed Detections
3.5 Proposed Architecture for Fusion of Models
4 Experiments and Results
4.1 Datasets
4.2 Metrics
5 Conclusion
References
Perceptual Image Anomaly Detection
1 Introduction
2 Related Work
3 Perceptual Image Anomaly Detection
3.1 Relative-perceptual-L1 Loss
3.2 Training Objective
3.3 Gradient-Normalizing Weight Policy
4 Experiments
4.1 Results
4.2 Ablation Study
5 Conclusion
References
Segmentation, Grouping and Shape
Deep Similarity Fusion Networks for One-Shot Semantic Segmentation
1 Introduction
2 Related Work
3 Problem Setup
4 Method
4.1 Dataset and Metric
4.2 Deep Similarity Fusion Network
4.3 Similarity Feature Fusion Module
5 Experiments
5.1 Implement Details
5.2 Results
5.3 Ablation Study
6 Conclusion
References
Seeing Things in Random-Dot Videos
1 Introduction
2 Signal Modelling
3 Literature Review
4 The A Contrario Framework
5 A Contrario on Random-Dot Videos
5.1 A Merging Strategy for Temporal Integration
5.2 Designing the A Contrario Algorithm
6 A Priori Analysis of A Contrario Performance
7 Empirical Results of the A Contrario Algorithm
7.1 Static Edge Case
7.2 Dynamic Edge Case
8 Human Performance Versus the A Contrario Process
8.1 Evaluating the Visual Angle
8.2 Evaluating the Time Integration
9 Concluding Remarks
References
Boundary Extraction of Planar Segments from Clouds of Unorganised Points
1 Introduction
2 Hough Voting Analysis
2.1 Hough Space
2.2 Voting Analysis
3 Plane Boundary Extraction
3.1 Top and Bottom Boundary Detection
3.2 Left and Right Boundary Detection
4 Experimental Results
4.1 Test on Synthetic Point Clouds
4.2 Test on Real-World Point Clouds
5 Conclusions
References
Real-Time Multi-class Instance Segmentation with One-Time Deep Embedding Clustering
1 Introduction
2 Related Works
3 One-Time Clustering Method
3.1 Encoder-Decoder Network
3.2 Loss Function
3.3 One-Time Clustering Algorithm
4 Experiments
4.1 Cityscapes Dataset
4.2 Evaluation Metrics
4.3 Training Environment Setup
4.4 Result and Discussion
5 Conclusion
References
Face and Body and Biometrics
Multi-person Pose Estimation with Mid-Points for Human Detection under Real-World Surveillance
1 Introduction
2 Methodology
2.1 Generating Body Region Points and Mid-Points
2.2 Generating Core-of-Pose
2.3 Generating Pose Vectors and Human Bounding Boxes
3 Evaluation
4 Discussion
4.1 Associating Parts Rather Than Detecting the Whole Target
4.2 Training Mid-Points
4.3 Triangles in Core-of-Pose
5 Conclusion and Future Work
References
Gaze from Head: Gaze Estimation Without Observing Eye
1 Introduction
2 Eye-Head Coordination
3 Methods
4 Datasets
4.1 Real Dataset
4.2 VR Dataset
5 Experiment
5.1 Qualitative Evaluation
5.2 Quantitative Evaluation
5.3 Compatibility of the Real and VR Datasets
6 Conclusion
References
Interaction Recognition Through Body Parts Relation Reasoning
1 Introduction
2 Related Work
2.1 Action Recognition
2.2 Human Interaction Recognition
2.3 Relational Network
3 Relational Network Overview
4 Interaction Relational Network
5 Experiments
5.1 Datasets
5.2 Implementation Details
5.3 Experimental Results
6 Conclusion
References
DeepHuMS: Deep Human Motion Signature for 3D Skeletal Sequences
1 Introduction
2 Related Work
3 DeepHuMS: Our Method
4 Experiments
4.1 Datasets
4.2 Implementation Details
4.3 Evaluation Metrics
4.4 Comparison with State of the Art
4.5 Discussion
4.6 Limitations
5 Conclusion
References
Adversarial Learning and Networks
Towards a Universal Appearance for Domain Generalization via Adversarial Learning
1 Introduction
2 Related Work
3 Proposed Method
3.1 Problem Statement and Motivations
3.2 Network Architecture
3.3 Appearance Generalizer
3.4 Adversarial Learning
4 Experiment
4.1 Evaluation of Appearance Generalizer
4.2 Comparison with Existing Methods
4.3 Ablation Study
5 Conclusion
References
Pre-trained and Shared Encoder in Cycle-Consistent Adversarial Networks to Improve Image Quality
1 Introduction
2 Related Work
2.1 Cycle-Consistent Adversarial Networks
2.2 Super-Resolution
3 Proposed Method
3.1 Encoder-Decoder Block
3.2 Partly Frozen CycleGAN
4 Experiments
4.1 Training Setting
4.2 Training Epoch of ED-Block
4.3 Results
4.4 Evaluation and Discussion
5 Conclusions
References
MobileGAN: Compact Network Architecture for Generative Adversarial Network
1 Introduction
2 Related Work
2.1 MobileNetV1: Separable Convolution
2.2 MobileNetV2: Inverted Residuals
3 Proposed Method
3.1 Compact Generator Network
3.2 Asymmetric GAN Architecture
4 Experiments
5 Conclusions
References
Denoising and Inpainting of Sea Surface Temperature Image with Adversarial Physical Model Loss
1 Introduction
2 Related Work
2.1 Data Assimilation
2.2 Image Restoration by Inpainting
3 Our Method
3.1 Loss Functions
3.2 Using Spatio-Temporal SST Images
3.3 Generator Function
3.4 Network Architecture
4 Experiments
5 Conclusion
References
Computational Photography
Confidence Map Based 3D Cost Aggregation with Multiple Minimum Spanning Trees for Stereo Matching
1 Introduction
2 Framework
3 Proposed Method
3.1 3D Cost Aggregation with Forest Structure
3.2 Reliable Disparity Extraction with Distinctiveness Analysis
3.3 Anchored Rolling Refinement
4 Experiments
4.1 Parameter Setting
4.2 Evaluation Based on Middlebury 3.0 Benchmark
4.3 Parameter Analysis
5 Conclusion
References
Optical Flow Assisted Monocular Visual Odometry
1 Introduction
2 Related Work
2.1 Supervised Approaches
2.2 Unsupervised Approaches
3 Methodology
3.1 Overview
3.2 Flow Prediction Part
3.3 Motion Estimation Part
3.4 Loss Function
4 Experiments
4.1 Datasets
4.2 Training Setup
4.3 How Our Flow Prediction Module Help VO Learning
4.4 VO Performances
5 Conclusions
References
Coarse-to-Fine Deep Orientation Estimator for Local Image Matching
1 Introduction
2 Related Work
2.1 Hand-Crafted Keypoint Detectors
2.2 CNN-Based Methods
3 Proposed Method
3.1 Coarse Orientation Estimator
3.2 Fine Orientation Estimator
3.3 The Calculation of Final Orientation
4 Experiments
4.1 Experimental Settings
4.2 Accuracies over Different Loss Functions
4.3 Evaluation of Area Under the Precision-Recall Curve
4.4 Comparison with Other Local Feature Descriptors
5 Conclusion
References
Geometric Total Variation for Image Vectorization, Zooming and Pixel Art Depixelizing
1 Introduction
2 Geometric Total Variation
3 Contour Regularization
4 Raster Image Zooming with Smooth Contours
5 Experimental Evaluation and Discussion
6 Conclusion
References
Learning Theory and Optimization
Information Theory-Based Curriculum Learning Factory to Optimize Training
1 Introduction
2 Related Work
3 Proposed Method
3.1 Stage I: Assessing Content of Training Samples
3.2 Stage II: Sorting Batches of Samples
3.3 Stage III: Syllabus Evaluation
4 Experiments
4.1 Datasets
4.2 Training
4.3 Networks
5 Results and Analysis
5.1 Training Trends
5.2 Regularization Loss
5.3 Sparsity
5.4 Classification Results
5.5 Conclusion
References
A Factorization Strategy for Tensor Robust PCA
1 Introduction
2 Notations and Preliminaries
3 TriFac for Tensor Robust PCA
3.1 Model Formulation
3.2 Optimization Algorithm
4 Experiments
4.1 Synthetic Data Experiments
4.2 Real Data Experiments
5 Conclusion
References
Speeding up of the Nelder-Mead Method by Data-Driven Speculative Execution
1 Introduction
2 Related Work
2.1 Problem Setting of Hyperparameter Optimization
2.2 Parallel HPO Methods
2.3 The Nelder-Mead (NM) Method
3 Speculative Execution for the NM Method
3.1 The Future Iterates of the NM Method
3.2 The Frequency of Each Operation
3.3 The Data-Driven Speculative NM Method
4 Experimental Evaluation
4.1 Experimental Conditions
4.2 The Comparison with Parallel HPO Methods
4.3 The Comparison with Other Parallel NM Methods
5 Conclusions
References
Efficient Bayesian Optimization Based on Parallel Sequential Random Embeddings
1 Introduction
2 Related Works
3 Settings and Random Embedding
4 Proposed Method
5 Experiment
5.1 Experimental Method
5.2 Results and Discussion
6 Conclusion
References
Applications, Medical and Robotics
Fundus Image Classification and Retinal Disease Localization with Limited Supervision
1 Introduction
2 Related Work
3 Method
3.1 Revisiting Grad-CAM
3.2 Heatmap Selection by Physician Intervention (PI)
3.3 Guidance with Attention Mining
4 Experiments
4.1 Datasets and Experimental Settings
4.2 Performance on Classification
4.3 Physician Intervention Is Necessary
4.4 Qualitative Results on Lesion Localization
5 Conclusion
References
OSTER: An Orientation Sensitive Scene Text Recognizer with CenterLine Rectification
1 Introduction
2 Related Works
3 Methodology
3.1 Centerline Rectification Network
3.2 Orientation Sensitive Network
3.3 Attention Selective Sequence-to-Sequence Network
3.4 Training
4 Experiments
4.1 Datasets Setup
4.2 Implementation Details
4.3 Performance of the ASN
4.4 Performances of the CRN
4.5 Performances on Public Datasets
5 Conclusion
References
On Fast Point Cloud Matching with Key Points and Parameter Tuning
1 Introduction
2 Related Work
3 Keypoint Selection
3.1 ISS3D
3.2 Harris3D
3.3 SIFT
4 Methodology
4.1 Keypoint Evaluation
4.2 Registration Pipeline
4.3 Parameter Tuning
5 Evaluation and Discussion
5.1 Data Sets
5.2 Results
6 Conclusion
References
Detection of Critical Camera Configurations for Structure from Motion Using Random Forest
1 Introduction
2 Related Work
2.1 Model Selection
2.2 Conjugate Rotation
2.3 Quality Measure
3 Classification
3.1 Features
3.2 Detection
4 Evaluation
4.1 Datasets
4.2 Analysis of Parameters
4.3 Analysis of Features
4.4 Integration into a Structure from Motion Framework
5 Conclusion
References
Computer Vision and Robot Vision
Visual Counting of Traffic Flow from a Car via Vehicle Detection and Motion Analysis
1 Introduction
2 Vehicle Detection in Video Frame via YOLO3
3 Motion Trace Generation and Vehicle Counting
3.1 Vehicles Motion Characteristics in Driving Video
3.2 Counting Vehicles Without Tracking Traces
4 Dense Vehicle Counting for Traffic Flow Estimation
5 Experiments
5.1 Naturalistic Driving Data for Testing
5.2 Counting Accuracy Along a Route
6 Conclusion
References
Directly Optimizing IoU for Bounding Box Localization
1 Introduction
2 Related Work
3 Methodology
3.1 Smooth IoU Loss
4 Experiments
4.1 Evaluation and Discussion
5 Conclusion
References
Double Refinement Network for Room Layout Estimation
1 Introduction
2 Related Works
3 Heat Map Estimation via Double Refinement Net
3.1 Network Architecture
4 Post-processing
4.1 Key Points Coordinate Estimation
4.2 Layout Ranking
5 Results
6 Conclusions
References
Finger Tracking Based Tabla Syllable Transcription
1 Introduction
2 The Instrument
2.1 The Instrument Bols
3 Literature Survey
4 Dataset Used
5 Proposed Methodology
5.1 Feature Extraction and Processing
5.2 Feature Classification
6 Results and Discussions
7 Conclusion and Future Work
References
Dual Templates Siamese Tracking
1 Introduction
2 Related Works
2.1 SiameseFC Networks
2.2 Difference Hash Algorithm
3 The Proposed Approach
3.1 Dual Templates Siamese Networks
3.2 Similarity Estimate
3.3 Adaptive Gaussian Background Suppression
3.4 Algorithm Flow
4 Experiments
4.1 OTB Datasets
4.2 VOT Datasets
5 Conclusion
References
Structure Function Based Transform Features for Behavior-Oriented Social Media Image Classification
1 Introduction
2 Related Work
3 Proposed Method
3.1 Foreground and Background Region Detection
3.2 Structural Function Based Transform (SFBT) for Feature Extraction
3.3 Behavior Oriented Social Media Image Classification
4 Experimental Results
4.1 Experiments for Two Class Classification
4.2 Experiments for Multi-class Classification
5 Conclusion and Future Work
References
Single Image Reflection Removal Based on GAN with Gradient Constraint
1 Introduction
2 Related Work
2.1 Optimization Based Methods
2.2 Deep Learning Based Methods
3 Supporting Methods
3.1 Synthetic Reflection Image
3.2 Generative Adversarial Networks
4 Proposed Method
4.1 Network Model
4.2 Loss Functions for Generator
4.3 Loss Function for Discriminator
4.4 Training Dataset
4.5 Training
4.6 Rotate Averaging Process
5 Experimental Results
5.1 Results
5.2 The Effectiveness of Our Loss Function and Training Method
6 Conclusion
References
SSA-GAN: End-to-End Time-Lapse Video Generation with Spatial Self-Attention
1 Introduction
2 Related Work
2.1 Generative Adversarial Networks
2.2 Video Generation
2.3 Video Prediction
2.4 Self Attention Mechanism
3 Our Approach
3.1 Spatial Self-Attention GAN
3.2 Spatial Self-Attention Module
3.3 Spatial Self-Attention GAN Objectives
4 Implementation Details
5 Experiments
5.1 Datasets
5.2 Experiments on the Cloud Time-Lapse Dataset
5.3 Experiments on the Beach Dataset
5.4 Discussion
6 Conclusion
References
Semantic Segmentation of Railway Images Considering Temporal Continuity
1 Introduction
2 Semantic Segmentation Considering Temporal Continuity
2.1 Overview of the Proposed Method
2.2 Semantic Segmentation Using Multiple Frames
3 Experimental Evaluation
3.1 Class Settings of the Railway Environment
3.2 Datasets
3.3 Experiment
4 Discussion and Applications
4.1 Improvement of Semantic Segmentation Accuracy
4.2 Effects of Using Class Likelihoods
4.3 Accuracy of Flow Estimation
4.4 Possible Applications to the Railway Environment
5 Conclusion
References
Real Full Binary Neural Network for Image Classification and Object Detection
1 Introduction
2 Related Work
2.1 Design Efficient Model
2.2 Quantized Neural Network
3 Real Full Binary Neural Network
3.1 Binary Weight Filters
3.2 Network Architecture
4 Experiments
4.1 Image Classification
4.2 Detection Result
5 Conclusion
References
Progressive Scale Expansion Network with Octave Convolution for Arbitrary Shape Scene Text Detection
1 Introduction
2 Related Work
2.1 Scene Text Detection
2.2 Methods for Enlarging Receptive Fields
3 Proposed Method
3.1 Overall Pipeline
3.2 Octave Convolution
3.3 Label Generation
3.4 Loss Function
3.5 Inference
4 Experimental Results
4.1 Datasets
4.2 Implementation Details
4.3 Comparisons with State-of-the-Art Methods
5 Conclusion
References
SFLNet: Direct Sports Field Localization via CNN-Based Regression
1 Introduction
2 Related Works
3 Proposed Approach
3.1 SFLNet Architecture
3.2 SFLNet Training
4 Experimental Evaluation
4.1 Dataset
4.2 Implementation Details
4.3 Ablation/Parameter Study
4.4 Comparison to Baselines
5 Conclusion
References
Trained Model Fusion for Object Detection Using Gating Network
1 Introduction
2 Related Work
2.1 Ensemble of Experts
2.2 Gating Network
2.3 Object Detection
3 Proposed Method
3.1 Overall Structure
3.2 Gating Network for Object Detection
3.3 Training Gating Network
4 Experiments
4.1 Setup
4.2 Results and Discussion
5 Conclusion
Appendices
A UA-DETRAC Dataset
B Accuracy of Trained Model on Test Dataset
References
Category Independent Object Transfiguration with Domain Aware GAN
1 Introduction
2 Related Work
3 Preliminaries
4 Domain Region and Magnitude Aware GAN
4.1 Architecture
4.2 Loss Functions
5 Implementation Details
6 Experiments
6.1 Few Object Categories Training Data
6.2 Wide Variety of Object Categories Training Data
7 Conclusion
References
New Moments Based Fuzzy Similarity Measure for Text Detection in Distorted Social Media Images
1 Introduction
2 Related Work
3 Proposed Method
3.1 Text Detection for Distorted Images
3.2 Moments Based Fuzzy Logic Similarity Measure for Text Detection
4 Experimental Results
4.1 Experiments on Distorted Social Media Images
4.2 Experiments on Benchmark Dataset of Natural Scene Images
5 Conclusion and Future Work
References
Fish Detection Using Convolutional Neural Networks with Limited Training Data
1 Introduction
2 Related Work
2.1 Background Subtraction
2.2 Convolutional Neural Network
3 Approach
3.1 Network Architecture
3.2 Data Augmentation Pre-processing
3.3 Activation Function and Loss Function
3.4 Post-processing
4 Experiments
4.1 Data Collection
4.2 Results
5 Conclusion
References
A New U-Net Based License Plate Enhancement Model in Night and Day Images
1 Introduction
2 Related Work
3 Proposed Model
3.1 The Proposed U-Net Based Enhancement Model
3.2 Text Detection for the Enhancement Images
4 Experimental Results
4.1 Evaluating the Proposed Enhancement Model
4.2 Validating the Proposed Enhancement Model Through Text Detection
5 Conclusion and Future Work
References
Integration of Biologically Inspired Pixel Saliency Estimation and IPDA Filters for Multi-target Tracking
1 Introduction
1.1 Target Detection via Bio-inspired Vision
1.2 Multi-target Tracking in Cluttered Environments
1.3 The Tracking Filter: Integrated Probabilistic Data Association
1.4 Statement of Contribution
2 Data Collection
3 Methodology
3.1 Target Extraction
3.2 Dropout Reduction via Linear Prediction
4 Experimental Validation
5 Conclusion
References
The Effectiveness of Noise in Data Augmentation for Fine-Grained Image Classification
1 Introduction
2 Related Work
3 Methods
3.1 Data Augmentation from the Web
3.2 NTS Network
3.3 Maximum Entropy Learning
4 Experiments
4.1 Datasets
4.2 Implementations
4.3 Results
4.4 Ablation Studies
5 Conclusion
References
Attention Recurrent Neural Networks for Image-Based Sequence Text Recognition
1 Introduction
2 Related Work
2.1 Attention Based RNNs
2.2 Attention Models for Image-Based Sequence Text Recognition
3 ARNNS
3.1 The Attention Gate
3.2 Attention Recurrent Neural Networks
4 ARNNs for Image-Based Sequence Text Recognition
5 Experiments and Results
5.1 House Number Recognition on the SVHN Dataset
5.2 Handwritten Word Recognition on the IAM Off-Line Dataset
5.3 Scene Text Recognition Based on the CRNN Architecture
6 Conclusion
References
Nighttime Haze Removal with Glow Decomposition Using GAN
1 Introduction
2 Preliminary Knowledge
2.1 Conditional Generative Adversarial Network
2.2 Atrous Convolution
3 Single Image Nighttime Haze Removal
3.1 Nighttime Haze Optical Model
3.2 Glow Removal with Generative Adversarial Network
3.3 Computation of Atmospheric Light
3.4 Estimation of Transmission Map
3.5 Scene Recovery
4 Experiment Results
4.1 Qualitative Analysis
4.2 Quantitative Analysis
5 Conclusion
References
Finding Logo and Seal in Historical Document Images - An Object Detection Based Approach
1 Introduction
2 Related Work
3 Motivation and Dataset
3.1 Data Aquisitition and Annotation
4 Methodology
4.1 Object Detection - Brief Overview
4.2 YOLOv3-A Fully Connected Network
4.3 Experimental Framework
5 Experimental Results and Discussions
5.1 YOLOv3 Network Learning Details and Obtained Results
5.2 Results on Tobacco800 Dataset
6 Conclusion and Future Work
References
KIL: Knowledge Interactiveness Learning for Joint Depth Estimation and Semantic Segmentation
1 Introduction
2 Related Work
3 Framework
3.1 Network Arichtecture
3.2 Knowledge Interactiveness Learning Module
3.3 Network Training
4 Experiment
4.1 Ablation Study
4.2 Comparisions with the State-of-the-Art Methods
5 Conclusion
References
Road Scene Risk Perception for Intelligent Vehicles Using End-to-End Affordance Learning and Visual Reasoning
1 Introduction
2 Related Work on Road Scene Understanding
2.1 Modular Methods
2.2 End-to-End Learning
2.3 Affordance Learning
2.4 Visual Reasoning
3 Proposed Method
3.1 Road Scene Risk Perception Using End-to-End Affordance Learning
3.2 Weakly Supervised Visual Reasoning of Underlying Risk Factors
4 Experimental Results and Evaluation
4.1 Using Saliency Maps for Visual Reasoning
4.2 Different Evaluation Strategy to Assess Safety and Efficiency of the Model
4.3 Comparison with Other Studies
5 Applications of Road Scene Risk Perception
6 Conclusion and Future Work
References
Image2Height: Self-height Estimation from a Single-Shot Image
1 Introduction
2 Related Work
2.1 Geometric Approach with Dynamic Scene Information
2.2 Recognition-Based Approach
2.3 Visualization Techniques for CNNs
3 Approach
3.1 Model
3.2 Fine Tuning and Preprocessing
3.3 Visualization and Sensitivity Analysis
4 Experiments
4.1 Data Collection
4.2 Accuracy Evaluation
4.3 Visualization of Important Regions
4.4 Sensitivity Analysis
5 Discussions
6 Conclusions
References
Segmentation of Foreground in Image Sequence with Foveated Vision Concept
1 Introduction
2 Foreground Detection Framework
3 Intensity and Motion Features
4 Photometric and Textural Features
5 Result
6 Conclusion
References
Bi-direction Feature Pyramid Temporal Action Detection Network
1 Introduction
2 Related Work
3 Methods
3.1 Network
3.2 Train
3.3 Inference
4 Experiments
4.1 Dataset and Setup
4.2 Ablation Study
4.3 Compare with State-of-the-Art
5 Conclusion
References
Hand Segmentation for Contactless Palmprint Recognition
1 Introduction
2 Related Work
2.1 U-Net ch64uspsnet
2.2 TernausNet ch64ternausnet
2.3 PSPNet ch64pspnet
2.4 DeepLab v3 ch64deeplabv3
3 Proposed Method
3.1 Encoder
3.2 PPM
3.3 Decoder
3.4 Loss Function for Hand Segmentation
4 Experiments and Discussion
5 Conclusion
References
Selecting Discriminative Features for Fine-Grained Visual Classification
1 Introduction
2 Related Work
2.1 Fine-Grained Visual Classification
2.2 Part Localization
3 Proposed Method
3.1 Feature Augmentation
3.2 Optimization
4 Experiments
4.1 Datasets
4.2 Implementation Details
4.3 Comparison with State-of-the-Art Methods
5 Conclusions
References
Author Index