Genereren van nieuwe aanzichten en inverse beeldsynthese voor glanzende voorwerpen: Efficiënte representaties, ambiguïteitsreductie en editeerbare modellen

Glossy objects and scenes are particularly difficult to reconstruct from images because their appearance changes strongly with viewpoint and lighting, and because reflections and highlights tightly entangle geometry, material properties, and illumination. This dissertation studies how to recover digital 3D representations of such scenes from posed images so that they can be rendered from novel viewpoints, relit under new illumination conditions, and edited in a controlled way. To achieve this, it develops efficient neural and Gaussian-based methods for novel view synthesis and inverse rendering, while also investigating the ambiguities that make glossy scene decomposition ill-posed and proposing ways to mitigate them. Overall, the work moves toward faster, more reliable, and more editable reconstructions of glossy scenes that are practical for downstream use in applications such as data augmentation, visual inspection, and graphics content creation.