Superpixel alpha-expansion and normal adjustment for stereo matching

Abstract

This paper presents a continuous stereo disparity estimation method based on superpixel segmentation and graph-cuts. We re-parameterize the disparity with a 3D tangent plane, and propose two algorithms to optimize the Markov Random Field (MRF) energy. The first algorithm, called superpixel α-expansion, is built on superpixel segmentation to localize the label proposal and the expansion scope. Three levels of superpixels with increasing granularity are generated for acceleration. The second algorithm, called normal adjustment, optimizes the 3D planes for the regions with low texture and/or illumination changes. The normal adjustment is performed along a depth-first similarity path of superpixels. We evaluate our method on the Middlebury 3.0 evaluation benchmark and the Eth3d benchmark. Experimental results show that our method achieves high accuracy on both evaluation benchmarks. (Middlebury 3.0 evaluation benchmark: http://vision.middlebury.edu/stereo/eval3/. Eth3d benchmark: https://www.eth3d.net/low_res_two_view.)