UC San Diego

High fidelity reconstruction and editing of objects is a challenging task in the graphics and vision community. Recent work in 3D reconstruction are unable to preserve high-frequency details in addition to enabling tasks such as texture transfer, primarily because they do not disentangle appearance from geometry. Further, reconstruction and editing methods for relighting applications learn a simplified reflectance model and are unable to account for long-range light transport effects such as subsurface scattering. This thesis presents two main directions of research for high fidelity reconstruction and object editing: First, we propose TEGLO (Textured EG3D-GLO) for learning textured 3D representations from single-view image collections. We train a conditional Neural Radiance Field (NeRF) without explicit 3D supervision and creating a dense correspondence mapping to a 2D canonical coordinate space to equip our method with texture transfer and editing with near perfect reconstruction (>74 db PSNR) even at megapixel resolution. Second, we find that recent work in high fidelity relighting explore subsurface scattering with objects where scattering is the primary light transport effect. These methods are unable to model specular highlights which occur when relighting human faces. Toward this, we render a synthetic OLAT dataset of human face images in a virtual light stage with suitable ground truth for reconstruction and relighting. We explore a hybrid physical-neural approach to surface relighting by effectively combining insights from a physically based prior and a neural renderer to improve the fidelity in modeling specular highlights and subsurface scattering effects in relighting human faces.