3D visual saliency detection model with generated disparity map

Abstract

Due to the remarkable distinction between human monocular vision and binocular vision, stereoscopic visual attention becomes an emerging question in the study of 3D applications. Some of existing 3D visual saliency detection models take advantage of ground-truth disparity map to compute center-surround differences of the depth features with high computational cost. In some 3D applications, the ground-truth disparity map may not be always available. In this paper, an efficient and simple 3D visual saliency detection model is proposed without using ground-truth disparity map. The proposed model is based on a band-pass filtering method which coincides with the visual perceptual process in human visual system. Firstly, the monocular luminance, color and texture features are extracted from the left view's image; the binocular depth feature is extracted from the two views' disparity map. Then, all the feature maps are filtered to generate three types of saliency maps, i.e., 2D saliency map, texture saliency map and depth saliency map. Subsequently, the three saliency maps are fused to one 3D saliency map by a linear pooling strategy. Finally, the final 3D visual saliency map is enhanced by the center-bias factor. Experimental results on a public eye tracking database show that the proposed model achieves better detection performance with low computational cost among the existing 3D visual saliency detection models.