Two-View Motion Segmentation by Gaussian Blurring Mean Shift with Fitness Measure

Motion segmentation for dynamic scene is fundamental in computer vision. The key issue is to estimate number and parameters of transformations simultaneously. However, transformations cannot be measured by general Euclidean distance because of them not lying in vector space. In this paper, we convert transformations into fitness vectors which can be easily measured by cosine similarity. Then we apply Gaussian blurring mean shift (GBMS) algorithm as a non-parameter clustering for further motion segmentation. Our approach doesn´t need any pre-knowledge of the number of motions and converges within a few iterations. Experimental results have shown that the method has a good performance for motion segmentation.