Region-based depth feature descriptor for saliency detection on light field

Abstract

This paper addresses the light field saliency detection problem via a multiple cue integration framework. By reinterpreting the usage of dark channels in estimating the amount of defocus, a novel region-based depth feature descriptor (RDFD) defined over the focal stack is proposed. Compared to the methods which utilize the depth map as another image channel, the RDFD can produce more informative saliency cues and make less restrictive assumptions on accurate depth map or focused clear images containing dark pixels. The proposed RDFD facilitates saliency detection in the following two respects: (1) the region-based depth contrast map can be computed by measuring a pair-wise distance between super-pixels with the proposed RDFD, (2) a spatial distribution prior in the 3D space (3D-SDP) can be obtained from such depth measurements to provide high-level semantic guidances, including the gradient-like distribution in depth and the object-biased prior in image plane. Both of them contribute to generating a contrast-based depth saliency map and refining a background-based color saliency map. Finally, these saliency maps are merged into a single map using a multi-layer cellular antomata (MCA) optimizer. Experimental results demonstrate that our method outperforms state-of-the-art techniques on the challenging light field saliency detection benchmark LFSD.