The aim of this paper is to analyze the discontinuity preserving behavior of two methods in optical flow. With this objetive, we have implemented a well-known optical flow method that uses isotropic TV-L 1 regularization. For the second approach, we have modified this method, by adding a decreasing function in the regularization term, to avoid smoothing at flow discontinuities. As a consequence, we see a high improvement and a very accurate discontinuities detection in some sequences but not good enough in others. Adapting the weight of the decreasing function allows us to better define the flow discontinuities. Nevertheless, the experimental results show that the parameters that yield a good segmentation of the motion field, may also introduce important unstabilities. In this sense, the results seem promising, but it is very difficult to set a unified parameter configuration that works fine for all the sequences. We evaluate the performance of these approaches with some standard test sequences, such as the Middlebury benchmark database or the Yosemite sequence. Looking for the best parameters configuration, which provides the best contour definition, does not typically mean a solution which is closer to the ground truth.

Full text: 81120117.pdf