Bootstrap Optical Flow Confidence and Uncertainty Measure

Abstract

We address the problem of estimating the uncertainty of optical flow algorithm results. Our method estimates the error magnitude at all points in the image. It can be used as a confidence measure. It is based on bootstrap resampling, which is a computational statistical inference technique based on repeating the optical flow calculation several times for different randomly chosen subsets of pixel contributions. As few as ten repetitions are enough to obtain useful estimates of geometrical and angular errors. For demonstration, we use the combined local-global optical flow method (CLG) which generalizes both Lucas-Kanade and Horn-Schunck type methods. However, the bootstrap method is very general and can be applied to almost any optical flow algorithm that can be formulated as a pixel-based minimization problem. We show experimentally on synthetic as well as real video sequences with known ground truth that the bootstrap method performs better than all other confidence measures tested.