Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021, Proceedings, Part I

Foreword by General Chairs
Preface
Challenges
ICPR Organization
Contents – Part I
3DHU 2020 - 3D Human Understanding
	Workshop on 3D Human Understanding (3DHU)
	Organization
	3DHU Workshop Chairs
	Technical Program Committee
	Additional Reviewer
	Sponsors
Subject Identification Across Large Expression Variations Using 3D Facial Landmarks
	1 Introduction
	2 Temporal Deformable Shape Model
	3 Experimental Design and Results
		3.1 3D Face Databases
		3.2 Experimental Design
		3.3 Subject Identification Results
		3.4 Subject Identification with Occluded Faces
		3.5 Comparisons to State of the Art
	4 Conclusion
	References
3D Human Pose Estimation Based on Multi-Input Multi-Output Convolutional Neural Network and Event Cameras: A Proof of Concept on the DHP19 Dataset
	1 Introduction
	2 Related Work
		2.1 HPE Datasets
		2.2 CNN Architectures
	3 Materials and Methods
		3.1 Dataset
		3.2 Preprocessing and Frame Generation
		3.3 Baseline: Single-Input Single-Output (SISO) Architecture
		3.4 Proposed Approach: Multiple-Input Multiple-Output (MIMO) Architecture
		3.5 3D Human Pose Estimation
		3.6 Experimental Procedure
	4 Results
		4.1 Validation Results
		4.2 2D Pose Estimation
		4.3 3D Pose Estimation
	5 Conclusions
	References
Image-Based Out-of-Distribution-Detector Principles on Graph-Based Input Data in Human Action Recognition
	1 Introduction
	2 Related Work
	3 Out-of-Distribution Detectors
		3.1 Baseline
		3.2 Out-of-DIstribution Detector for Neural Networks
		3.3 Learning Confidence for OoD Detection
		3.4 Metric Learning-Based Approach
	4 Evaluation
		4.1 Pipeline
		4.2 Semi-synthetic Dataset
		4.3 Metrics
		4.4 Experimental Setup
		4.5 Results
	5 Conclusion
	References
A Novel Joint Points and Silhouette-Based Method to Estimate 3D Human Pose and Shape
	1 Introduction
	2 Related Work
	3 Method
		3.1 Parametric Human Body Model
		3.2 Pose Fitting
		3.3 Shape Fitting
		3.4 Optimization
	4 Experiments
		4.1 Datasets
		4.2 Evaluation of Pose Fitting and Shape Fitting
		4.3 Comparison to Previous Approaches
	5 Conclusion
	References
Pose Based Trajectory Forecast of Vulnerable Road Users Using Recurrent Neural Networks
	1 Introduction
		1.1 Motivation
		1.2 Related Work
		1.3 Main Contributions and Outline of This Article
	2 Method
		2.1 Data Acquisition and Preprocessing
		2.2 Network Architecture
		2.3 Evaluation Method
	3 Experimental Results
		3.1 Hyperparameter Optimization
		3.2 Test Results
	4 Conclusions and Future Work
	References
Towards Generalization of 3D Human Pose Estimation in the Wild
	1 Introduction
	2 Related Datasets
	3 Proposed 3DBodyTex.Pose Dataset
	4 Experimental Evaluation
		4.1 Baseline 3D Pose Estimation Approach
		4.2 Data Augmentation with 3DBodyTex.Pose
	5 Conclusion
	References
Space-Time Triplet Loss Network for Dynamic 3D Face Verification
	1 Introduction
	2 Proposed Static 3D Face Representation
	3 Dynamic 3D Face Embedding
	4 Experimental Results
		4.1 Dataset and Experimental Setup
		4.2 Face Verification Results
	5 Conclusions and Future Works
	References
AIDP - Artificial Intelligence for Digital Pathology
	Preface
	Organization
	General Chairs
	Program Committee Chair
	Program Committee
Noise Robust Training of Segmentation Model Using Knowledge Distillation
	1 Introduction
	2 Related Work
		2.1 Noise Robust Training
		2.2 Knowledge Distillation
	3 Method
	4 Experiments
		4.1 Datasets
		4.2 Implementation Details
		4.3 Evaluation
	5 Conclusion
	References
Semi-supervised Learning with a Teacher-Student Paradigm for Histopathology Classification: A Resource to Face Data Heterogeneity and Lack of Local Annotations
	1 Introduction
	2 Methods
		2.1 Datasets
		2.2 Teacher/Student Paradigm
		2.3 Implementation
	3 Results
	4 Discussion
	5 Conclusion
	References
Self-attentive Adversarial Stain Normalization
	1 Introduction
	2 Related Work
	3 Approach
	4 Dataset and Implementation
		4.1 Dataset
		4.2 Network Architecture
		4.3 Training Details
	5 Results and Evaluation
	A  Conclusions
	B  Additional Results
	References
Certainty Pooling for Multiple Instance Learning
	1 Introduction
	2 Related Work
	3 Proposed Method
	4 Experiments
		4.1 Low Evidence Ratio MNIST-Bags
		4.2 Camelyon16 Lymph Node Metastasis Detection Challenge
	5 Conclusions
	References
Classification of Noisy Free-Text Prostate Cancer Pathology Reports Using Natural Language Processing
	1 Introduction
	2 Related Work
	3 Methods
		3.1 Corpus
		3.2 Corpus Preprocessing
		3.3 Data Augmentation
		3.4 Document Representation
		3.5 Document Classification
		3.6 Experimental Setup
	4 Results
	5 Discussion and Analysis
	6 Conclusions and Future Work
	References
AI Slipping on Tiles: Data Leakage in Digital Pathology
	1 Introduction
	2 Data Description
	3 Methods
	4 Results
	5 Discussion
	6 Conclusions
	References
AIHA 2020 - International Workshop on Artificial Intelligence for Healthcare Applications
	Preface
	Organization
	AIHA Chairs
	Program Committee
	Additional Reviewers
Predictive Medicine Using Interpretable Recurrent Neural Networks
	1 Introduction
	2 Related Work
	3 Background
		3.1 Recurrent Neural Networks (RNN)
		3.2 Long Short-Term Memory (LSTM)
		3.3 LSTM with Varying Timestamps
		3.4 SHAP Values
	4 Methodology
		4.1 Data
		4.2 Data Utils Package
		4.3 Interpretability
		4.4 Dashboard
		4.5 Reproducibility
	5 Results
		5.1 Model Performance
		5.2 Model Interpretation
	6 Conclusions
	References
Automated Detection of Adverse Drug Events from Older Patients' Electronic Medical Records Using Text Mining
	1 Introduction
	2 Methods
		2.1 De-Identification
		2.2 Manual Annotation
		2.3 Automatic Annotation
		2.4 OGER and BioBERT
		2.5 BioBERT
		2.6 Harmonisation and Merging
		2.7 Drug Administration
	3 Results
		3.1 De-Identification
		3.2 Lexico-Semantic Resources
		3.3 Automatic Annotation
	4 Discussion
	References
Length of Stay Prediction for Northern Italy COVID-19 Patients Based on Lab Tests and X-Ray Data
	1 Introduction
	2 Related Work
	3 Available Data Sources
		3.1 Data Quality Issues
	4 Datasets for Training and Testing
		4.1 Pre-processing and Feature Extraction
		4.2 Training and Test Sets Generation
	5 Machine Learning Algorithms
		5.1 Regression Algorithms
		5.2 Hyperparameter Search
	6 Experimental Evaluation and Discussion
		6.1 Results
	7 Conclusions and Future Work
	References
Advanced Non-linear Generative Model with a Deep Classifier for Immunotherapy Outcome Prediction: A Bladder Cancer Case Study
	1 Introduction
	2 Related Works
	3 The Proposed Deep Network Framework
		3.1 The Bounding Box Segmentation Block
		3.2 The 2D-CNN Features Generative Model
		3.3 The 2D-DNN Classifier with Decision System
	4 Experimental Results
	5 Conclusion and Discussion
	References
Multi-model Ensemble to Classify Acute Lymphoblastic Leukemia in Blood Smear Images
	1 Introduction
	2 Prior Art
	3 Materials and Methods
		3.1 Dataset
		3.2 Methodology
	4 Results and Discussion
	5 Conclusion
	References
MIINet: An Image Quality Improvement Framework for Supporting Medical Diagnosis
	1 Introduction
	2 Proposed Method – MIINet
		2.1 The Image Dehazing Module – IDM
		2.2 The Image Super-Resolution Module - ISR
	3 Experimental Results
		3.1 Throat Image Dataset
		3.2 Training the IDM
		3.3 Training the ISR Module
		3.4 The Mean Doctor Opinion Score
		3.5 Results
	4 Discussion
	5 Conclusion
	References
Medical Image Tampering Detection: A New Dataset and Baseline
	1 Introduction
	2 Tampered Medical Image Dataset Generation
	3 Framework for Medical Tampering Detection
		3.1 Architectures
		3.2 Model Parameterization
	4 Experiments and Results
		4.1 Experimental Setup
		4.2 Ablation Study of Backbone Networks
		4.3 Results of Our Framework
		4.4 Generalizability of the ConnectionNet
	5 Conclusion
	References
Deep Learning for Human Embryo Classification at the Cleavage Stage (Day 3)
	1 Introduction
	2 Method
		2.1 Images
		2.2 State-of-the-Art CNN Model: STORK
		2.3 Our CNN Models
		2.4 Implantation Rate
	3 Results
		3.1 Regression Model of Technician Scores for Standalone Cases
		3.2 STORK Performance on Middle Slice and Multi-slice Images
		3.3 CNN Performance on Middle Slice and Multi-slice Images in Standalone Fashion
		3.4 CNN Performance on Individual Technician Scores
		3.5 Batch Effect
		3.6 Performance on Easy Decisions
		3.7 Pregnancy Outcomes
	4 Discussion and Conclusion
	References
Double Encoder-Decoder Networks for Gastrointestinal Polyp Segmentation
	1 Introduction
	2 Methodology
		2.1 Double Autoencoders
		2.2 Pretrained Encoders
		2.3 Decoders
		2.4 Training Details
	3 Experimental Results
		3.1 Data and Evaluation Metrics
		3.2 Autoencoders vs. Double Autoencoders
		3.3 Comparison with Recent Techniques
		3.4 Qualitative Analysis
	4 Discussion and Conclusion
	References
A Superpixel-Wise Fully Convolutional Neural Network Approach for Diabetic Foot Ulcer Tissue Classification
	1 Introduction
	2 Related Work
	3 Proposed Method
		3.1 Image Acquisition and Data Annotation
		3.2 Ulcer Segmentation and Superpixels Extraction
		3.3 Superpixel-Based Tissue Classification
	4 Results
		4.1 Performance Metrics
		4.2 Experimental Results
	5 Conclusion
	References
Fully vs. Weakly Supervised Caries Localization in Smartphone Images with CNNs
	1 Introduction
		1.1 Medical Motivation
		1.2 Technical Motivation
		1.3 Contributions
	2 Related Work
	3 Methods
		3.1 Fully Supervised Object Detection
		3.2 Weakly Supervised Localization
		3.3 Implementation
	4 Experiments
		4.1 Data
		4.2 Evaluation
	5 Discussion
	6 Conclusion
	References
Organ Segmentation with Recursive Data Augmentation for Deep Models
	1 Introduction
	2 Dataset
	3 Proposed Methodology
	4 Experimental Setup
	5 Results and Discussion
	6 Conclusion
	References
Pollen Grain Microscopic Image Classification Using an Ensemble of Fine-Tuned Deep Convolutional Neural Networks
	1 Introduction
	2 Materials and Methods
		2.1 Dataset
		2.2 Pre-processing
		2.3 Pre-trained CNNs
		2.4 Fine-Tuning
		2.5 Fusion
		2.6 Evaluation
		2.7 Implementation
	3 Results
	4 Discussion
	5 Conclusions
	References
Active Surface for Fully 3D Automatic Segmentation
	1 Introduction
	2 Materials and Methods
		2.1 Patient Dataset and PET Protocol Acquisition
		2.2 Overview of the Proposed System
		2.3 Performance Evaluation
	3 Results
	4 Discussions
	References
Penalizing Small Errors Using an Adaptive Logarithmic Loss
	1 Introduction
	2 Adaptive Logarithmic Loss
	3 Evaluation
	4 Results
	5 Conclusion
	References
Exploiting Saliency in Attention Based Convolutional Neural Network for Classification of Vertical Root Fractures
	1 Introduction
	2 Relate Work
		2.1 VRFs Recognition
		2.2 CNN Based Image Classification
		2.3 Weakly Supervised Learning
	3 Materials and Method
		3.1 VRFs DataSets
		3.2 Feature Pyramids Attention Convolutional Neural Network
	4 Experiments
	References
UIP-Net: A Decoder-Encoder CNN for the Detection and Quantification of Usual Interstitial Pneumoniae Pattern in Lung CT Scan Images
	1 Introduction
	2 State of the Art
		2.1 Deep Learning and Convolutional Neural Networks
	3 Data and Methods
		3.1 Data
		3.2 Methods
	4 Experimental Setup and Results
		4.1 Experimental Setup
		4.2 Results
		4.3 Discussion
	5 Conclusion
	References
Don’t Tear Your Hair Out: Analysis of the Impact of Skin Hair on the Diagnosis of Microscopic Skin Lesions
	1 Introduction
	2 Related Work for Hair Detection
	3 Methods
		3.1 Hair Segmentation
		3.2 Augmentations
		3.3 Skin Lesion Classification
	4 Results
		4.1 Hair Segmentation
		4.2 Skin Lesion Classification
	5 Conclusions and Future Work
	References
Deep Learning Based Segmentation of Breast Lesions in DCE-MRI
	1 Introduction
	2 State of the Art
	3 Materials and Methods
		3.1 Data
		3.2 Pre-processing
		3.3 Segmentation Algorithm
	4 Results and Discussion
		4.1 Experiment 1: Threshold
		4.2 Experiment 2: Optimizer
		4.3 Experiment 3: Loss Function
		4.4 Experiment 4: Patch Size
		4.5 Implementation Details
		4.6 Discussion
	5 Conclusions and Future Works
	References
Fall Detection and Recognition from Egocentric Visual Data: A Case Study
	1 Introduction
	2 Related Works
		2.1 Fall Detection by Fixed Visual Sensors
		2.2 Fall Detection by Wearable Devices
		2.3 Fall Detection by Wearable Cameras
	3 Data Set
	4 Method
		4.1 Extraction of Frames
		4.2 Feature Extraction
		4.3 Fusion of Features and Classification Model
	5 Experiments and Results
		5.1 Experimental Design
		5.2 Performance Measurements
		5.3 Results
	6 Discussion
	7 Conclusions
	References
Deep Attention Based Semi-supervised 2D-Pose Estimation for Surgical Instruments
	1 Introduction
	2 Related Work
	3 Methodology
		3.1 Network Architecture
		3.2 Post-processing
		3.3 Total Variation as a Confidence Measure for Pose Estimation
		3.4 Training Details
	4 Experiments
		4.1 Datasets
		4.2 Results Using RMIT Dataset
		4.3 Results Using Endovis Dataset
	5 Conclusion
	References
Development of an Augmented Reality System Based on Marker Tracking for Robotic Assisted Minimally Invasive Spine Surgery
	1 Introduction
	2 Methods
		2.1 Marker System Selection
		2.2 Server-Client Communication
		2.3 Evaluation Protocol
	3 Results
		3.1 Localization Accuracy of the Pose Estimation
		3.2 Runtime
		3.3 Robustness to External Influences
		3.4 Communication and 3D Visualization
	4 Discussion
	5 Conclusion
	References
Towards Stroke Patients' Upper-Limb Automatic Motor Assessment Using Smartwatches
	1 Introduction
	2 State of the Art
	3 Experimentation Protocol
	4 Methodology
		4.1 Data Capture and Preprocessing
		4.2 Data Labelling
		4.3 Segmentation
		4.4 Classification
	5 Results
		5.1 Dataset
		5.2 Results
	6 Conclusion
	References
Multimodal Detection of Tonic–Clonic Seizures Based on 3D Acceleration and Heart Rate Data from an In-Ear Sensor
	1 Introduction
	2 Related Work
	3 Method
		3.1 Data Preparation
		3.2 Learning Process
	4 Experiments
		4.1 Dataset
		4.2 Experimental Setup
		4.3 Results
		4.4 Discussion
	5 Conclusion and Future Work
	References
Deep Learning Detection of Cardiac Akinesis in Echocardiograms
	1 Introduction
	2 Related Works
	3 Proposed Method
		3.1 Dataset
		3.2 Network Specifications
	4 Experimental Results
	5 Conclusions
	References
Prediction of Minimally Conscious State Responder Patients to Non-invasive Brain Stimulation Using Machine Learning Algorithms
	1 Introduction
		1.1 Methods
	2 Results
		2.1 Pre-EEG (Stimulation and Sham)
		2.2 Pre-EEG (Stimulation and Sham) Augmented with Bootstrap
		2.3 Discussion
	3 Conclusions
	References
Sinc-Based Convolutional Neural Networks for EEG-BCI-Based Motor Imagery Classification
	1 Introduction
	2 Sinc Layer
	3 The Sinc-EEGNet Architecture
	4 Experiments
	5 Results
	6 Conclusions
	References
An Analysis of Tasks and Features for Neuro-Degenerative Disease Assessment by Handwriting
	1 Introduction
	2 Classic Velocity-Based Features
	3 Additional Kinematic-Based Features
		3.1 Maxwell-Boltzmann Distribution
		3.2 Discrete Transformations
	4 Experiment
	5 Conclusions
	References
A Comparative Study on Autism Spectrum Disorder Detection via 3D Convolutional Neural Networks
	1 Introduction
	2 Related Work
		2.1 ASD Detection
		2.2 3D Convolutional Neural Networks (CNN)
	3 3D CNN for ASD
		3.1 C3D
		3.2 I3D
		3.3 3D ResNet
		3.4 The Proposed 3D ResNeSt
		3.5 Dataset
		3.6 Data Preprocessing
	4 Experiments and Results
		4.1 Implementation Details
		4.2 Experimental Results
	5 Conclusion
	References
A Multi Classifier Approach for Supporting Alzheimer's Diagnosis Based on Handwriting Analysis
	1 Introduction
	2 Related Work
	3 The Multi Classifier Architecture
	4 Experimental Results
		4.1 Implementation Details
		4.2 Dataset
		4.3 Tasks Characterisation
		4.4 Baseline Evaluation Session
		4.5 The Basic Classifiers
		4.6 Combining All
		4.7 Combining the Best
	5 Conclusions
	References
A Lightweight Spatial Attention Module with Adaptive Receptive Fields in 3D Convolutional Neural Network for Alzheimer’s Disease Classification
	1 Introduction
	2 Related Work
		2.1 Deep Learning Methods for AD Classification
		2.2 Attention Mechanism
		2.3 Dilated Convolution
	3 Method
		3.1 3D Spatial Attention Module with Adaptive Receptive Fields
		3.2 Data and Preprocessing
		3.3 Experiment Setup
	4 Result and Discussion
		4.1 Comparisons Using Different Single-Branch Cases
		4.2 Comparisons Using Different Two-Branch Cases
		4.3 Comparisons Using Different Single-Branch and Two-Branch Cases
		4.4 Comparisons with Related Studies
	5 Conclusion
	References
Handwriting-Based Classifier Combination for Cognitive Impairment Prediction
	1 Introduction
	2 The Tasks
	3 Data Collection and Feature Extraction
	4 The Proposed Approach
	5 Experiments and Results
	6 Conclusions
	References
CADL2020 - Workshop on Computational Aspects of Deep Learning
	Preface
	Organization
	General Chairs
	Program Committee
WaveTF: A Fast 2D Wavelet Transform for Machine Learning in Keras
	1 Introduction
	2 Background
		2.1 Wavelet Transform
		2.2 TensorFlow and Keras
	3 Related Work
		3.1 PyWavelets
		3.2 pypwt
		3.3 TF-Wavelets
	4 Implementation
		4.1 Direct Transform
		4.2 Inverse Transform
		4.3 Correctness
	5 Performance Results
		5.1 Raw Transformation
		5.2 Machine Learning
	6 Software Availability
	7 Conclusion and Future Work
	References
Convergence Dynamics of Generative Adversarial Networks: The Dual Metric Flows
	1 Introduction
		1.1 Motivation: W-GANs
		1.2 Mathematical Setting
	2 Theoretical Results
		2.1 Motivation and Literature Review
		2.2 Basic Reminders
		2.3 Definition of (EDI Style) Equilibrium Flows
		2.4 Convergence of Numerical Schemes
	3 Applications
	4 Discussion and Conclusion
	References
Biomedical Named Entity Recognition at Scale
	1 Introduction
	2 NER Model Implementation in Spark NLP
	3 Implementation Details and Experimental Results
		3.1 Datasets
		3.2 Overview of Experimental Setup
		3.3 Experiment Results
	4 Conclusion
	A  Appendices
	References
PyraD-DCNN: A Fully Convolutional Neural Network to Replace BLSTM in Offline Text Recognition Systems
	1 Introduction
	2 Related Work
	3 PyraD-DCNN Model
		3.1 Design Principles
		3.2 Overall Architecture
	4 Experiments
		4.1 Data
		4.2 Implementation Details
		4.3 Ablation Study
	5 Results
		5.1 Experiment on Small-ID
		5.2 Experiment on Big-ID
		5.3 Discussion
	6 Conclusion
	References
Learning Sparse Filters in Deep Convolutional Neural Networks with a l1/l2 Pseudo-Norm
	1 Introduction
	2 Related Work
		2.1 Network Pruning
		2.2 Weight Sparsity
	3 Training with Kernel-Sparsity
		3.1 Kernel-Sparsity Regularization
		3.2 Training with Kernel-Sparsity Regularization
		3.3 Setting Kernels to Zero
	4 Experiments
		4.1 Experiments on LeNet
		4.2 VGG on CIFAR10
	5 Conclusion
	References
Multi-node Training for StyleGAN2
	1 Introduction
	2 Multi-node Training via Horovod
		2.1 Process Parallelism
		2.2 Data Sharding
		2.3 Gradient Averaging
		2.4 Multi-node Metrics
	3 Validation
	4 Scaling Tests
		4.1 Strong Scaling
		4.2 Weak Scaling
	5 Conclusion
	References
Flow R-CNN: Flow-Enhanced Object Detection
	1 Introduction
	2 Related Work
		2.1 Region-Based Methods
		2.2 Regression-Based Methods
		2.3 Flow-Based Object Detection
	3 Flow R-CNN
		3.1 Object-Based Motion Analysis
		3.2 Mask R-CNN
		3.3 Proposed Architecture
	4 Experimental Results
		4.1 Object Detection Datasets
		4.2 Experimental Environment
		4.3 Comparative Evaluation
	5 Conclusions
	References
Compressed Video Action Recognition Using Motion Vector Representation
	1 Introduction
	2 Related Works
	3 Proposed Approach
		3.1 Motion Vector
		3.2 Key Information Selection
		3.3 Motion Vector Representation
		3.4 Baseline Model
	4 Experiments
		4.1 Datasets and Experimental Details
		4.2 Accuracy and Efficiency
		4.3 Ablation Studies
		4.4 Visualizations
	5 Conclusions
	References
Introducing Region Pooling Learning
	1 Introduction
	2 Related Work
	3 Region Pooling Learning
	4 Experiments and Results
		4.1 Average-Max Pooling Behavior
		4.2 CIFAR-10
		4.3 ImageNet
	5 Conclusion
	References
Second Order Bifurcating Methodology for Neural Network Training and Topology Optimization
	1 Introduction
	2 Related Work
	3 Method
		3.1 Horizontal Tangent Parabola (HTP)
		3.2 Vertical Tangent Parabola (VTP)
		3.3 Algorithm
		3.4 Example 1
		3.5 Example 2
	4 Training a Radial Basis Function NN
	5 Experiments
		5.1 Interpolation of a 2D Surface
		5.2 Learning the Kernel Surface
		5.3 Application in Convolutional Neural Networks
	6 Conclusions
	References
Author Index