Interactive Stereo Image Segmentation with RGB-D Hybrid Constraints

Wei Ma,Yue Qin,Luwei Yang,Shibiao Xu,Xiaopeng Zhang
DOI: https://doi.org/10.1109/lsp.2016.2605133
2016-01-01
IEEE Signal Processing Letters
Abstract:This letter presents an approach to extracting a target object interactively from a given pair of stereo images. First, a user marks a few parts of the object and background in either of the two views with strokes. The marked pixels are used to generate the prior models of the foreground and background. Second, a graph is constructed with constraints formulated by the priors of foreground/background, similarities between intraview neighbor pixels and correspondences between interview pixels. Third, two segments of the foreground are extracted from the two views by optimization of the graph via graph cut. Traditional methods generally define the priors and neighbor similarities in RGB space. Differently, the proposed method integrates disparity distributions of foreground/background to enrich the priors and defines the similarity metric between neighbor pixels in RGB-D space. The proposed method that utilizes RGB-D hybrid constraints generates stereo segments with accuracies higher than those obtained by state-of-the-art methods.
What problem does this paper attempt to address?