Master's Thesis: Reliable Deep Surface Normal Estimation through Probabilistic Quantification of Aleatoric Uncertainty from Monocular Images
This repository hosts the research and experimental work conducted for my Master's thesis, which builds upon "Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation". It includes additional experiments, enhancements, and personal insights on the original CNN-based approach for surface normal estimation in the context of 3D scene understanding.
My Master's thesis explores advanced methodologies in surface normal estimation from single images, particularly emphasizing the aleatoric uncertainty aspects in 3D scene understanding. This repository documents my journey in enhancing and experimenting with the original CNN approach proposed by Bae et al., aiming to refine the capture of aleatoric uncertainty and detail in predictions.
- To investigate and expand upon the original CNN architecture for surface normal estimation.
- To incorporate novel experimental strategies for improving uncertainty estimation.
- To analyze the impact of these enhancements on prediction quality, especially in complex scene geometries.
- Experimentation with comparative novel loss functions