Interactive segmentation method on stereo images
H. Emrah Tasli, A. Aydın Alatan · 2012
Segmentation problem dates back to a couple of decades and many solutions have been proposed so far. The fact that there is still no solution with optimum performance justifies the efforts towards new studies. This study concentrates on keeping the user interaction in the loop for superior results. Proposed technique extends previous knowledge on monoscopic image segmentation to stereoscopic footage with little extra effort. Image representative regions corresponding to background and foreground are indicated through human supervision. An MRF based energy minimization technique where user inputs are applied only on one of the stereoscopic pairs is utilized. Propagation of single view information on the stereo pair is supplied by a sparse feature matching idea. This eliminates the necessity of dense disparity matching and hence removes a computationally complex step from the pipeline. Segmentation results have been evaluated on a ground truth stereo segmentation dataset and superior results with minimum user interaction are obtained even without dense disparity estimation.