Non-Local smoothness constraints for disparity estimation in a variational framework

The Non-Local Total Variation (NLTV) has been recently formalized to define new functionals for signal and image analysis, that strictly fit into the widely used variational framework but overcome the locality limitation of the classical TV. This work lies in the context of disparity estimation in a variational framework, where Total Variation represents a common tool to impose a smooth behavior to the desired solution. Here, with reference to a recently proposed disparity estimation technique, several new smoothness constraints based on a NLTV formulation are presented, to prove the effectiveness of the non-local approach in encompassing structural prior knowledge in the problem. Results on several stereo pairs from the Middlebury database are very encouraging, and highlight the importance of a more accurate formulation of the smoothness constraint in the disparity estimation problem.