Postprocessing of optical flows via surface measures and motion inpainting

Dense optical flow fields are required for many applications. They can be obtained by means of various global methods which employ regularization techniques for propagating estimates to regions with insufficient information. However, incorrect flow estimates are propagated as well. We, therefore, propose surface measures for the detection of locations where the full flow can be estimated reliably, that is in the absence of occlusions, intensity changes, severe noise, transparent structures, aperture problems and homogeneous regions. In this way we obtain sparse, but reliable motion fields with lower angular errors. By subsequent application of a basic motion inpainting technique to such sparsified flow fields we obtain dense fields with smaller angular errors than obtained by the original combined local global (CLG) method and the structure tensor method in all test sequences. Experiments show that this postprocessing method makes error improvements of up to 38% feasible.

Lecture Notes in Computer Science