Pattern Recognition: 45th DAGM German Conference, DAGM GCPR 2023, Heidelberg, Germany, September 19–22, 2023, Proceedings (Lecture Notes in Computer Science, 14264)

This book constitutes the proceedings of the 45th Annual Conference of the German Association for Pattern Recognition, DAGM-GCPR 2023, which took place in Heidelberg, Germany, during September 19-22, 2023. The 40 full papers included in these proceedings were carefully reviewed and selected from 76 submissions. They were organized in topical sections as follows: Segmentation and action recognition; 3D reconstruction and neural rendering; Photogrammetry and remote sensing; Pattern recognition in the life sciences; Interpretable machine learning; Weak supervision and online learning; Robust models. Preface Organization Contents Segmentation and Action Recognition Score-Based Generative Models for Medical Image Segmentation Using Signed Distance Functions 1 Introduction 2 Method 2.1 Image Segmentation Using SDFs 2.2 Conditional Score-Based Segmentation 2.3 Motivation of the SDF in Conditional Score-Based Segmentation 3 Experiments 3.1 Data Sets 3.2 Architecture and Training 3.3 Sampling 3.4 Evaluation 3.5 Results 3.6 Segmentation Uncertainty 3.7 Ablation 4 Discussion and Limitations 5 Conclusion and Outlook References A Trimodal Dataset: RGB, Thermal, and Depth for Human Segmentation and Temporal Action Detection 1 Introduction 2 Related Work 2.1 Datasets 2.2 Methods 3 System Overview 3.1 Sensor Setup 3.2 Ground Truth Generation 4 Trimodal Dataset 4.1 Dataset Design 4.2 Dataset Analysis 4.3 Dataset Quality Evaluation 5 Benchmarking 5.1 Split 5.2 Human Segmentation 5.3 Action Detection 6 Conclusion References Airborne-Shadow: Towards Fine-Grained Shadow Detection in Aerial Imagery 1 Introduction 2 Shadow Detection Datasets 3 Airborne-Shadow Dataset 3.1 Shadow Annotation 3.2 Comparison to the Other Datasets 4 Shadow Detection Methods 5 Evaluation Metrics 6 Results and Discussion 7 Conclusion and Future Works References UGainS: Uncertainty Guided Anomaly Instance Segmentation 1 Introduction 2 Related Work 2.1 Pixel-Wise Anomaly Segmentation 2.2 Anomaly Instance Segmentation 2.3 Promptable Segmentation Models 3 Method 4 Experiments 4.1 Results 4.2 Ablations 5 Conclusion References Local Spherical Harmonics Improve Skeleton-Based Hand Action Recognition 1 Introduction 2 Related Work 2.1 Skeleton-Based Action Recognition 2.2 Frequency Domain Representations for Action Recognition 2.3 Spherical Harmonics in 3D Point Clouds 3 Method 3.1 Local Spherical Coordinates 3.2 Spherical Harmonics Based Hand Joint Representations 3.3 Models 3.4 Evaluation 4 Experiments 4.1 Datasets 4.2 Implementation Details 4.3 Experimental Results 5 Discussion and Conclusion References 3D Reconstruction and Neural Rendering LMD: Light-Weight Prediction Quality Estimation for Object Detection in Lidar Point Clouds 1 Introduction 2 Related Work 3 Proposed Method 4 Numerical Results 5 Conclusion References A Network Analysis for Correspondence Learning via Linearly-Embedded Functions 1 Introduction 2 Related Work 2.1 Non-rigid Correspondence Methods 2.2 Functional Map-Based Learning Approaches 3 Background 3.1 Functional Maps 3.2 Linearly-Invariant Embedding 4 Method 4.1 Unsupervised Training 4.2 Subsampling Scheme 5 Experiments 5.1 Datasets 5.2 Evaluation 5.3 Correspondence Accuracy 5.4 Different Sampling 5.5 Ablation Study 6 Conclusion References HiFiHR: Enhancing 3D Hand Reconstruction from a Single Image via High-Fidelity Texture 1 Introduction 2 Related Work 2.1 Parametric Hand Models 2.2 Hand Texture Reconstruction 3 Methods 3.1 Overview 3.2 Model Structure 3.3 Training Objective 4 Experiment 4.1 Implementation Details 4.2 Datasets and Evaluation Metrics 4.3 Comparison to State-of-the-Art Methods 4.4 Evaluation on Texture Reconstruction Consistency with Different Supervision Settings 4.5 Ablation Study 5 Conclusions References Point2Vec for Self-supervised Representation Learning on Point Clouds 1 Introduction 2 Related Work 3 Method 3.1 Data2vec 3.2 Data2vec for Point Clouds 3.3 Point2vec 4 Experiments 4.1 Self-supervised Pre-training 4.2 Main Results on Downstream Tasks 4.3 Analysis 5 Conclusion References FullFormer: Generating Shapes Inside Shapes 1 Introduction 2 Related Work 3 Method 3.1 Sequential Encoding with VQUDF 3.2 Generating a Sequence of Latent Vectors 4 Experiments 4.1 Implementation Details 4.2 VQUDF Reconstruction Performance 4.3 Baselines 4.4 Metrics 4.5 Qualitative Results 4.6 Quantitative Results 4.7 Limitations 5 Conclusion References GenLayNeRF: Generalizable Layered Representations with 3D Model Alignment for Human View Synthesis 1 Introduction 2 Related Work 2.1 Neural View Synthesis 2.2 Human Mesh Recovery 3 Methodology 3.1 Problem Definition 3.2 Human-Image Alignment Module 3.3 Layered Scene Representation 3.4 Feature Extraction and Attention-Aware Fusion 3.5 Radiance Field Predictor 3.6 Layered Volumteric Rendering and Loss Functions 4 Experiments 4.1 Datasets 4.2 Baselines 4.3 Experimental Results 4.4 Ablation Studies 5 Limitations and Future Work 6 Conclusion References RC-BEVFusion: A Plug-In Module for Radar-Camera Bird's Eye View Feature Fusion 1 Introduction 2 Related Work 2.1 Image-Only Object Detection 2.2 Radar-Only Object Detection 2.3 Sensor Fusion Object Detection 3 RC-BEVFusion 3.1 Overview 3.2 Radar Encoders 3.3 Camera-Only Baselines 3.4 Detection Head and Loss 4 Experimental Results 4.1 Data and Metrics 4.2 Quantitative Evaluation 4.3 Qualitative Evaluation 5 Conclusion References Parallax-Aware Image Stitching Based on Homographic Decomposition 1 Introduction 2 Related Work 3 Parallax-Aware Image Stitching Pipeline 3.1 Dense Feature Matching 3.2 Overlapping Region 3.3 Non-overlapping Region 3.4 Stitching 4 Experiments and Discussion 4.1 Evaluation 4.2 Results 5 Conclusions References Photogrammetry and Remote Sensing DustNet: Attention to Dust 1 Introduction 2 Related Work 2.1 Vision Transformer 2.2 Crowd Counting 2.3 Monocular Depth Estimation 3 Method 3.1 Network Structure 3.2 Temporal Fusion 4 Experimental Results 4.1 Dataset 4.2 Benchmark Selection 4.3 Implementation Details 4.4 Evaluation Metrics 4.5 Quantitative Results 4.6 Qualitative Results 4.7 Ablation Study 4.8 Limitations and Future Work 5 Conclusion References Leveraging Bioclimatic Context for Supervised and Self-supervised Land Cover Classification 1 Introduction 2 Related Work 3 Methodology 3.1 Data Preparation 3.2 Neural Network Architectures and Training 3.3 Leveraging Bioclimatic Data 4 Experiments and Evaluation 4.1 Accuracy 4.2 Generalizability 4.3 Training Efficiency 4.4 Autoencoder Reconstructions and Representations 4.5 Sanity Checks for Conditional Batch Normalization 4.6 Towards Large-Scale Application 5 Conclusion References Automatic Reverse Engineering: Creating Computer-Aided Design (CAD) Models from Multi-view Images 1 Introduction 2 Related Work 2.1 Traditional Photogrammetry Approaches to Reconstructing CAD Models 2.2 Learning-Based Object Reconstruction 2.3 Multi-view Convolutional Networks 2.4 Recurrent Convolutional Networks 2.5 Generation of CAD Representations 3 Methods 3.1 Network Architecture 3.2 Two-Stage Training 4 Experimental Setup 4.1 Training Data 4.2 Hyper-parameter Optimization 4.3 Accuracy Metrics 4.4 Benchmark Comparison 4.5 Reconstruction from Photographic Images 5 Results 6 Discussion and Conclusions References Characterization of Out-of-distribution Samples from Uncertainty Maps Using Supervised Machine Learning 1 Introduction 1.1 Motivation 1.2 Problem Statement 1.3 Our Solution in a Nutshell 2 Related Work 3 Methodology 3.1 Uncertainty Determination 3.2 Histogram Analysis 3.3 Connected Component Analysis 3.4 Evaluation 4 Experimental Setup 4.1 Dataset Preparation 4.2 Deep Learning Models 4.3 Evaluation Strategy 5 Results and Discussion 5.1 Semantic Segmentation 5.2 Uncertainty Analysis 5.3 Connected Component Analysis 6 Conclusion and Future Work References Underwater Multiview Stereo Using Axial Camera Models 1 Introduction and Previous Work 2 Axial Model with Virtual Camera Approximation 3 Refractive PatchMatch 4 Evaluation 5 Conclusion References Pattern Recognition in the Life Sciences 3D Retinal Vessel Segmentation in OCTA Volumes: Annotated Dataset MORE3D and Hybrid U-Net with Flattening Transformation 1 Introduction 2 2D vs. 3D Vessel Detection 3 Flattening Transformation for 3D OCTA Images 4 OCTA Dataset MORE3D with 3D Labeled Vessel Network 5 3D Vascular Network Segmentation 5.1 Vesselness Measures 5.2 Hybrid U-Net Architecture 5.3 Flattening-Based Data Augmentation 6 Experimental Results 6.1 Evaluation Metrics 6.2 Vesselness Measures on OCTA Images 6.3 3D Vessel Segmentation in OCTA Images 7 Conclusion References M(otion)-Mode Based Prediction of Ejection Fraction Using Echocardiograms 1 Introduction 2 Methods 2.1 From B-Mode Videos to M-Mode Images 2.2 Learning Representations from M-Mode Images 3 Experiments and Results 3.1 Dataset 3.2 Experimental Setup 3.3 Results and Discussion 4 Discussion and Conclusion References Improving Data Efficiency for Plant Cover Prediction with Label Interpolation and Monte-Carlo Cropping 1 Introduction 2 Related Work 3 Methods 3.1 Label Interpolation 3.2 Monte-Carlo Cropping 4 Experimental Results 4.1 Dataset 4.2 Setup 4.3 Metrics 4.4 Label Interpolation 4.5 Monte-Carlo Cropping 5 Conclusion and Future Work References Learning Channel Importance for High Content Imaging with Interpretable Deep Input Channel Mixing 1 Introduction 2 Related Work 3 Model 3.1 Conception of the Image Blending Layer 3.2 Classifying Genetic Perturbations Based on DCMIX-Blended Images 3.3 End-to-End Training Algorithm 4 Experiments 4.1 MNIST 4.2 RXRX1 5 Discussion 6 Conclusion References Self-supervised Learning in Histopathology: New Perspectives for Prostate Cancer Grading 1 Introduction 2 Materials and Methods 2.1 Self-supervised Training Methods 2.2 Dataset and Training Details 3 Experiments 3.1 Qualitative Analysis 3.2 Patch-Wise Performance Analysis 3.3 WSI-Wise Performance Analysis 4 Discussion References Interpretable Machine Learning DeViL: Decoding Vision features into Language 1 Introduction 2 Related Work 3 DeViL Model 3.1 Translating Vision Features into Language 3.2 Learning to Decode Individual Vision Features 3.3 Generalization Through Dropout 3.4 Open-Vocabulary Saliency with DeViL 4 Experiments 4.1 Evaluating Feature Descriptions Through Image Captioning 4.2 MILAN: Explaining Local Visual Features 4.3 Diverse Layer-Wise Explanations of Vision Features 4.4 Inspecting Different Vision Models Through Saliency 5 Conclusion References Zero-Shot Translation of Attention Patterns in VQA Models to Natural Language 1 Introduction 2 Related Work 3 The ZS-A2T Framework 4 Experiments 4.1 Comparing to Related Frameworks 4.2 Ablation Studies on Guiding Inputs 4.3 Language Models 4.4 Prompt Ablations 4.5 Qualitative Results 5 Limitations 6 Conclusion References Beyond Debiasing: Actively Steering Feature Selection via Loss Regularization 1 Introduction 2 Related Work 3 Method 3.1 Feature Steering 3.2 Feature Attribution 3.3 Theoretical Considerations 4 Datasets 4.1 Redundant Regression Dataset 4.2 Colored MNIST 5 Experiments 5.1 Evaluation Metrics 5.2 Results on Redundant Regression Dataset 5.3 Results on Colored MNIST 6 Conclusions References Simplified Concrete Dropout - Improving the Generation of Attribution Masks for Fine-Grained Classification 1 Introduction 2 Related Work 3 Simplified Concrete Dropout - Improved Stability 3.1 The FIDO Algorithm and Its Limitations 3.2 Improving Computational Stability 3.3 Combined Attribution Mask for Fine-Grained Classification 4 Experiments 4.1 Evaluating Mask Precision 4.2 Mask Coherency 4.3 Test-Time Augmentation of a Fine-Grained Classifier 5 Conclusions References Weak Supervision and Online Learning Best Practices in Active Learning for Semantic Segmentation 1 Introduction 2 Deep Active Learning 2.1 Active Learning for Semantic Segmentation 2.2 Semi-supervised Active Learning 3 Experimental Setup 3.1 Tested Approaches 3.2 Datasets 3.3 Experiment Details 4 Results 4.1 Impact of Dataset Redundancy 4.2 Systematic Integration of SSL 4.3 Low Annotation Budget 4.4 An Exemplar Case Study: A2D2-3K Task 5 Conclusion References COOLer: Class-Incremental Learning for Appearance-Based Multiple Object Tracking 1 Introduction 2 Related Work 3 Method 3.1 Problem Definition 3.2 COOLer 3.3 Continual Pseudo-label Generation for Tracking 3.4 Class-Incremental Instance Representation Learning 4 Evaluation Protocol 5 Experiments 5.1 Baselines 5.2 Implementation Details 5.3 Experimental Results 5.4 Ablation Study 6 Conclusion References Label Smarter, Not Harder: CleverLabel for Faster Annotation of Ambiguous Image Classification with Higher Quality 1 Introduction 1.1 Related Work 2 Methods 2.1 Simulated Proposal Acceptance 2.2 CleverLabel 2.3 Implementation Details 3 Evaluation 3.1 Label Improvement 3.2 Benchmark Evaluation 4 Discussion 5 Conclusion References Speeding Up Online Self-Supervised Learning by Exploiting Its Limitations 1 Introduction 2 Related Works 3 Problem Definition 4 Proposed Method 4.1 Single Local Cycle Learning Rate Scheduler 4.2 Reducible Anchor Loss Selection 4.3 Integration of SSL Methods 5 Experimental Results 5.1 Datasets 5.2 Settings 5.3 Implementation Details 5.4 CIFAR-100 Results 5.5 ImageNet-100 Results 5.6 Analysis 6 Conclusion References Text-to-Feature Diffusion for Audio-Visual Few-Shot Learning 1 Introduction 2 Related Work 3 Audio-Visual (G)FSL Benchmark 3.1 Audio-Visual (G)FSL Setting 3.2 Dataset Splits and Training Protocol 3.3 Benchmark Comparisons 4 AV-Diff Framework 4.1 Audio-Visual Fusion with Cross-Modal Attention 4.2 Text-Conditioned Feature Generation 4.3 Training Curriculum and Evaluation 5 Experiments 5.1 Implementation Details 5.2 Audio-Visual GFSL Performance 5.3 AV-Diff Model Ablations 6 Conclusion References Correlation Clustering of Bird Sounds 1 Introduction 2 Related Work 3 Model 3.1 Representation of Clusterings 3.2 Bayesian Model 4 Learning 5 Inference 6 Experiments 6.1 Dataset 6.2 Metrics 6.3 Clustering vs Classification 6.4 Clustering Unseen Data 6.5 Clustering Noise 7 Conclusion References MargCTGAN: A ``Marginally'' Better CTGAN for the Low Sample Regime 1 Introduction 2 Related Works 2.1 Tabular Data Generators 2.2 Evaluation of Tabular Data Generators 3 Background 3.1 Tabular GAN (TableGAN) 3.2 Conditional Tabular GAN (CTGAN) 3.3 Tabular VAE Model (TVAE) 4 Method: MargCTGAN 5 Multi-dimensional Evaluation Metrics 6 Implementation Details 6.1 Metrics 7 Experiments 7.1 Correlation of Metrics 7.2 Performance Comparison 8 Discussion 9 Conclusion References Robust Models Detecting Model Misspecification in Amortized Bayesian Inference with Neural Networks 1 Introduction 2 Related Work 3 Method 3.1 Neural Posterior Estimation 3.2 Model Misspecification in Simulation-Based Inference 3.3 Structured Summary Statistics 3.4 Detecting Model Misspecification with Finite Data 4 Experiments 4.1 Experiment 1: 2D Gaussian Means 4.2 Experiment 2: Cancer and Stromal Cell Model 4.3 Experiment 3: Epidemiological Model for COVID-19 5 Conclusions References Adversarial Perturbations Straight on JPEG Coefficients 1 Introduction 2 Related Work 2.1 Adversarial Attacks 2.2 Perceptual Metrics 2.3 JPEG Compression and JPEG-Resistant Attacks 2.4 Adversarial Defenses 2.5 Adversarial Attacks and Defenses from a Frequency Perspective 3 Proposed Method 4 Experiments and Results 4.1 Varying Luma and Chroma Perturbations 4.2 Varying Perturbations Across Frequencies 4.3 Comparison of Adversarial Attacks on JPEG Coefficients to YCbCr and RGB Pixel Representations 4.4 Comparison with JPEG-Resistant Attacks 5 Conclusion References Certified Robust Models with Slack Control and Large Lipschitz Constants 1 Introduction 2 Related Work 3 Calibrated Lipschitz-Margin Loss (CLL) 3.1 Background 3.2 Binary CLL 3.3 Discussion 4 Evaluation 4.1 Two-Moons Dataset 4.2 Image Datasets 4.3 Analysis and Ablation 5 Conclusion References Multiclass Alignment of Confidence and Certainty for Network Calibration 1 Introduction 2 Related Work 3 Proposed Methodology 3.1 Definition and Quantification of Calibration 3.2 Proposed Auxiliary Loss: MACC 4 Experiments 5 Conclusion References Drawing the Same Bounding Box Twice? Coping Noisy Annotations in Object Detection with Repeated Labels 1 Introduction 2 Related Work 3 Method 3.1 Localization-Aware Expectation-Maximization 3.2 Algorithmic Design Choices 3.3 Comparative Analysis 4 Experimental Results 4.1 Set-Up 4.2 Annotation Budget Ablation 4.3 Leave-One-Out Annotator Selection 5 Conclusion References An Evaluation of Zero-Cost Proxies - from Neural Architecture Performance Prediction to Model Robustness 1 Introduction 2 Related Work 2.1 Zero-Cost Proxies for NAS 2.2 Robustness in NAS 3 Background on Zero-Cost Proxies 3.1 Jacobian-Based 3.2 Pruning-Based 3.3 Piecewise Linear 3.4 Hessian-Based 3.5 Baselines 4 Feature Collection and Evaluation 4.1 NAS-Bench-201 4.2 Neural Architecture Design and Robustness Dataset 4.3 Collection 5 Evaluations of Zero-Cost Proxies 5.1 Correlation 5.2 Feature Importance 5.3 Feature Importance Excluding Top 1 6 Conclusion References Author Index