Methods and applications of uncertainty quantification for object recognition / submitted by Tobias Riedlinger, M.Sc. Wuppertal, September 4, 2023
Inhalt
- Acknowledgements
- Foreword
- Introduction
- Foundations
- Statistical Learning Theory
- Deep Learning
- Architectures
- Training Neural Networks
- Universal Approximation
- Single-Layer Perceptrons
- Perceptrons with ReLU Activation
- Convolutional Neural Networks
- Transformer Neural Networks
- Uncertainty Quantification
- Advanced Computer Vision Tasks
- Gradient Uncertainty for Deep Object Detectors
- Introduction
- Related Work
- Method
- Experiments
- Implementation Details
- Comparison with Output-based Uncertainty
- Generalization over Architectures
- Calibration
- Pedestrian Detection
- MetaFusion
- Runtime
- Conclusion
- Gradient Uncertainty in Semantic Segmentation
- Introduction
- Related Work
- Methods
- Experiments
- Implementation Details
- Pixel-wise Uncertainty Quantification
- Segment-wise Uncertainty Quantification
- Out-of-Distribution Segmentation
- Runtime
- Conclusion
- Rapid Prototyping of Active Learning for Object Detection
- Introduction
- Related Work
- Methods
- Experiments
- Implementation Details
- Benchmark Results
- Generalization of Sandbox Results
- Image-Aggregation Methods
- Runtime
- Conclusion
- Label Error Identification For Object Detection Datasets
- Prediction Quality Estimation for Lidar Object Detection
- Introduction
- Related Work
- Methods
- Experiments
- Implementation Details
- Correlation of Features and IoUBEV
- Meta Classification Models
- Generalization over Datasets and Networks
- Feature Selection
- Confidence Calibration
- Label Error Detection
- Conclusion
- Conclusion and Outlook
- List of Notations