Learning visual saliency for stereoscopic images

摘要

Currently, there are various saliency detection models proposed for saliency prediction in 2D images/video in the previous decades. With the rapid development of stereoscopic display techniques, stereoscopic saliency detection is much desired for the emerging stereoscopic applications. Compared with 2D saliency detection, the depth factor has to be considered in stereoscopic saliency detection. Inspired by the wide applications of machine learning techniques in 2D saliency detection, we propose to use the machine learning technique for stereoscopic saliency detection in this paper. The contrast features from color, luminance and texture in 2D images are adopted in the proposed framework. For the depth factor, we consider both the depth contrast and depth degree in the proposed learned model. Additionally, the center-bias factor is also used as an input feature for learning the model. Experimental results on a recent large-scale eye tracking database show the better performance of the proposed model over other existing ones.