Saliency detection for stereoscopic video

Abstract

We present a novel system for automatically detecting salient image regions in stereoscopic videos. Our proposed algorithm considers information based on three dimensions: salient colors in individual frames, salient information derived from camera and object motion, and depth saliency. These three components are dynamically combined into one final saliency map based on the reliability of the individual saliency detectors. Such a combination allows using more efficient algorithms even if the quality of one detector degrades. For example, we use a computationally efficient stereo correspondence algorithm that might cause noisy disparity maps for certain scenarios. In this case, however, a more reliable saliency detection algorithm such as the image saliency is preferred. To evaluate the quality of the saliency detection, we created modified versions of stereoscopic videos with the non-salient regions blurred. Having users rate the quality of these videos, the results show that most users do not detect the blurred regions and that the automatic saliency detection is very reliable.