A Novel Object-Oriented Stereo Matching on Multi-scale Superpixels for Low-Resolution Depth Mapping

This paper presents a novel object-oriented stereo matching on multi-scale superpixels to generate a low-resolution depth map. It overcomes the classic downsampling methods' disadvantages, such as boundary blurring, outlier enlargement and foreground objects merging to background, etc. The approach we exploited is to segment the image in three scales' superpixels from dense to sparse ones according to downsampling scale first, then compute disparity directly on superpixel's stereo matching. The post-processing of region constraint and local refinement uses hierarchical multi-scale superpixels as well. The proposed approach is validated on Middle-bury test-bed, and the experimental results outperform the current state-of-the-art stereo matching methods in low resolutions.