Determining Generative Models of Objects Under Varying Illumination: Shape and Albedo from Multiple Images Using SVDand Integrability

摘要: We describe a method of learning generative models objects from set images the object under different, and unknown, illumination. Such model allows us to approximate objects‘ appearance range lighting conditions. This work is closely related photometric stereo with unknown light sources and, in particular, use Singular Value Decomposition (SVD) estimate shape albedo multiple up linear transformation (Hayakawa, 1994). Firstly we analyze extend SVD approach this problem. demonstrate that it applies for which dominant imaging effects are Lambertian reflectance distant source background ambient term. To determine reasonable approximation calculate eigenvectors on real objects, varying conditions, first few account most data agreement our predictions. then ambiguities previous methods proposed resolve them 1994) only valid certain discuss alternative possibilities knowledge class sufficient Secondly, surface consistency putting constraints possible solutions. prove constraint reduces subspace called generalized bas relief ambiguity (GBR) inherent function (and can be shown exist even if attached cast shadows present (Belhumeur et al., 1997)). solve GBR describe, implement, variety additional assumptions GBR. Thirdly, an iterative algorithm detect remove some thereby increasing accuracy reconstructed albedo.