Visual recognition using Deep Neural Networks on abstract and decomposed data / submitted by Annika Mütze, M.Sc. Wuppertal, 29.02.2024
Inhalt
- Acknowledgments
- Foreword
- Introduction
- Fundamentals
- Deep Neural Networks
- Learning from Data
- Statistical Learning
- Error Decomposition
- Universal Approximation
- Training Process
- Training Stabilization by Normalization and Regularization Layers
- Deep Learning for Images
- Overcoming Data Limitation
- Generative Learning
- Semi-supervised Task Aware Image-to-Image Domain Adaptation
- Introduction
- Related Work
- Semi-supervised Task Aware I2I Translation
- Stage a) – Training the Task Expert
- Stage b) – Unsupervised Image-to-Image Translation
- Stage c) – Downstream Task Awareness
- Complete SSDA Method
- Sampling Strategies for the Labeled Subset TR
- Evaluation
- Conclusion and Outlook
- Cooperation Is All You Need?
- Introduction
- Related Work
- Cue Decomposition
- Late Fusion
- Cue Influence Analysis
- Base Datasets
- Implementation
- Experimental Setup and Evaluation Metrics
- Cityscapes Experiments
- CARLA Experiments
- Domain Shift Due to Cue Reduction
- Influence on Class Level
- Influence on the per Pixel Level
- Influence of the Architecture: Transformer Experiments
- Likelihood Based Evaluation
- Conclusion, Discussion and Outlook
- Discussion and Outlook
- List of Notations