Local smoothness for global optical flow

We consider the problem of estimating the "smoothness parameter" that controls the tradeoff between data fidelity and regularity in optical flow estimation. We start by reviewing the problem of global estimation using the Optimal Prediction Principle (OPP) by Zimmer et al. [1]. Inspired by this technique and work on local-global optical flow we propose a simple method for fusing optical flow estimates of different smoothness by evaluating interpolation quality locally by means of L
¹
block match on the corresponding set of gradient images. We illustrate the method in a setting where optical flows are estimated by a TV-L
¹
energy. On average this procedure reduces the average endpoint error by 15% over flows estimated using the OPP, and gives flow fields that are consistently better than the single best flows with a fixed smoothness parameter.