Evaluation and Improvement of Region-Based Motion Segmentation

Several approaches of motion segmentation were published in the last years, but an evaluation of these different approaches is missing up to now. Here we evaluate different methods of motion segmentation to optimize our motion estimation system [5] based on a n:m matching of color regions. We compare four different neighborhood checking methods and three different motion similarity tests in combination. For this purpose a quality measure was developed with a hand segmentation as a ground truth. This measure contains both the positive motion segmentation error and the negative one. Moreover a new, efficient approach for checking neighborhood relations between image regions is presented and also evaluated.