Local Similarity Pattern and Cost Self-Reassembling for Deep Stereo Matching Networks

Biyang Liu, Huimin Yu

Abstract:

Although convolution neural network based stereo matching architectures have made impressive achievements, there are
still some limitations: 1) The appearance information captured by Convolutional Feature (CF) is inadequate for accurate matching and susceptible to illumination change. 2) Due to the static mechanism, current disparity refinement methods often produce over-smooth results. In this paper, we present two schemes to respectively address these issues. Firstly, we introduce a pairwise feature LSP (Local Similarity Pattern).Through explicitly revealing the neighbor relationships, LSP contains rich structural information. On the one hand, LSP
can aid CF for more discriminative description. On the other hand, it is more robust to illumination and color change. Secondly,
we propose a dynamic self-reassembling refinement strategy and respectively apply it to the cost distribution and the disparity map. The former can be equipped with the unimodal distribution constraint to alleviate the over-smoothing problem, and the latter is more practical. Extensive experiments will demonstrate the effectivenesses of our methods. Specifically, via incorporating the designed modules, we surpass the basic model GwcNet by a significant margin on SceneFlow and KITTI benchmarks. We also insert them into GANet-deep and achieve a performance improvement.