High-resolution depth maps can be inferred from low-resolution depthmeasurements and an additional high-resolution intensity image of the samescene. To that end, we introduce a bimodal co-sparse analysis model, which isable to capture the interdependency of registered intensity and depthinformation. This model is based on the assumption that the co-supports ofcorresponding bimodal image structures are aligned when computed by a suitablepair of analysis operators. No analytic form of such operators exist and wepropose a method for learning them from a set of registered training signals.This learning process is done offline and returns a bimodal analysis operatorthat is universally applicable to natural scenes. We use this to exploit thebimodal co-sparse analysis model as a prior for solving inverse problems, whichleads to an efficient algorithm for depth map super-resolution.
展开▼
