Superpixel-based graph cuts for accurate stereo matching

Abstract

Abstract. Estimating the surface normal vector and disparity of a pixel simultaneously, also known as three-dimensional label method, has been widely used in recent continuous stereo matching problem to achieve sub-pixel accuracy. However, due to the infinite label space, it's extremely hard to assign each pixel an appropriate label. In this paper, we present an accurate and efficient algorithm, integrating patchmatch with graph cuts, to approach this critical computational problem. Besides, to get robust and precise matching cost, we use a convolutional neural network to learn a similarity measure on small image patches. Compared with other MRF related methods, our method has several advantages: its sub-modular property ensures a subproblem optimality which is easy to perform in parallel; graph cuts can simultaneously update multiple pixels, avoiding local minima caused by sequential optimizers like belief propagation; it uses segmentation results for better local expansion move; local propagation and randomization can easily generate the initial solution without using external methods. Middlebury experiments show that our method can get higher accuracy than other MRF-based algorithms.

A Review on Stereo Vision Algorithm: Challenges and Solutions

“…Moreover, the Patchmatch algorithm was integrated with GC by Feng, et al [71] to reduce computational complexity. Using the Patchmatch method, the label of pixels is randomly selected and propagated to adjacent pixels.…”

Multi-Path Minimum Spanning Tree and Superpixel Based Cost Aggregation for Stereo Matching

“…The global algorithms usually don't need cost aggregation. Popular global methods are dynamic programming [10], [11], belief propagation [12], [15], [16], [31] and graph cuts [13], [30]. For local algorithms, the disparity calculation at a given image pixel only relies on intensity or color values within a specific window.…”