Shape from video: Dense shape, texture, motion and lighting from monocular image streams.

摘要

We present a probabilistic framework for robust recovery of dense 3D shape, motion, texture and lighting from monocular image streams. We assume that the object is rigid, smooth, Lambertian, illuminated by one distant light source and subject to transformations that are smoothly time-varying. The problem is formulated as a large optimization where we learn all model and pdf (probability distribution function) parameters simultaneously, using a quasi-Newtonian optimization technique.