Physics aware learning of intrinsic images

This thesis explores physics-aware learning of intrinsic images; reflectance (albedo) and illumination (shading). The main goal is devoted to steering the learning processes of deep convolutional neural networks by physics-based reflection models and physics-based invariant descriptors. We demonstrate how to combine existing laws into the design of machine learning models. Besides that, semantic segmentation, optical flow, and surface normal modalities are investigated and the correlations between them are also explored. Findings of this thesis may lead to enhanced computer vision applications that are more robust to illumination variations and photometric effects.