Stereo Matching Based on Density Segmentation and Non-Local Cost Aggregation.

Recently, segment-tree based Non-Local cost aggregation algorithm, which can provide extremely low computational complexity and outstanding performance, has been proposed for stereo matching. The segment-tree (ST) based method integrated the segmentation information with non-local cost aggregation. However, the segmentation method used in the ST method results in under-segmentation so that some of the edges crossing the boundary will be preserved. On the other hand, pixel-level color information can not represent different patterns (smooth regions, texture and boundaries) well. So, only using the color information to establish the weight function is not enough. We proposed a density information based ST for non-local cost aggregation method. The core idea of the algorithm includes: (1) SLIC based method via density information is used to segment image. This clustering feature (density feature) and over-segmentation method are more suitable for stereo matching. (2) In generating sub-MST and linking all the sub-MSTs, we use density information to establish the weight function. We not only consider the color information but also the density information when establishing the weight formula. Performance evaluations on 31 Middlebury stereo pairs show the proposed algorithm outperforms better than other state-of-the-art aggregated based algorithms.