Non-convex optimization for robust multi-view imaging

We study the multi-view imaging problem where one has to reconstruct a set of l images, representing a single scene, from a few measurements made at different viewpoints. We first express the solution of the problem as the minimizer of a non-convex objective function where one needs to estimate one reference image, l foreground images modeling possible occlusions, and a set of l transformation parameters modeling the inter-correlation between the observations. Then, we propose an alternating descent method that attempts to minimize this objective function and produces a sequence converging to one of its critical points. Finally, experiments show that the method accurately recovers the original images and is robust to occlusions.