Oriented-linear-tree based cost aggregation for stereo matching

Matching cost aggregation is one of the most important steps in dense stereo correspondence, and non-local cost aggregation methods based on tree structures have been widely studied recently. In this paper, we analyze the shortcomings of both the local window-based aggregation methods and the non-local tree-based aggregation methods, and propose a novel oriented linear tree structure for each pixel to perform the non-local cost aggregation strategy. Firstly, each pixel in the image has an oriented linear tree rooted on it and each oriented linear tree consists of multiple 1D paths from different directions. Compared to other spanning trees, our oriented linear trees don’t need to be additionally constructed beforeh and since they are naturally embedded in the original image. Moreover, each root pixel not only gets supports from adjacent pixels within its local support window, but also receives supports from the other pixels along all 1D paths. Secondly, for each pixel lying on the same 1D path, we can at the same time calculate their aggregated cost along their path by traversing the path back and forth twice. Finally, the final aggregated cost for each root pixel can be calculated by summing the aggregated costs from all 1D paths. Performance evaluation on the Middlebury and KITTI datasets shows that the proposed method outperforms the current state-of-the-art aggregation methods.