Local PatchMatch Based on Superpixel Cut for Efficient High-resolution Stereo Matching

Abstract

Obtaining the accurate disparity of each pixel quickly is the goal of stereo matching, but it is very difficult for the 3D labels-based methods due to huge search space of 3D labels, especially for highresolution images. We present an novel two-stage optimization strategy to get the accurate disparity map for highresolution stereo image efficiently, which includes feature points optimization and superpixel optimization. In the first stage, we construct the support points including edge points and robust points for triangulation, which is used to extract feature points and then perform spatial propagation and random refinement to get the candidate 3D label sets. In the stage of superpixel optimization, we update per pixel labels of the corresponding superpixels using the candidate label sets, and then perform spatial propagation and random refinement. In order to provide more prior information to identify weak texture and textureless areas, we design the weight combination of "intensity + gradient + binary image" for constructing an optimal minimum spanning tree (MST) to compute the aggregated matching cost, and obtain the labels of minimum aggregated HIGHLIGHTS• A new method of extracting the feature points in superpixels is proposed to compute the candidate disparities using non-local cost aggregation.• A novel two-stage optimization framework for superpixel proposals by first getting the labels of feature points quickly and then consists the candidate label sets for updating the labels of pixels within the corresponding superpixel instead of assigning a label randomly to those pixels.• The weight combination of intensity, gradient and binary image is designed for constructing an optimal minimum spanning tree to compute the aggregated matching cost.

A Semi-Supervised Method for PatchMatch Multi-View Stereo with Sparse Points

“…Previous semi-supervised domains are all trained on the basis of MVS and not implemented using the PatchMatch method [3][4][5]. This is because with MVS, the depth map is obtained by constructing the cost volume, and we only need to refine the depth map, while in the PatchMatch method, the depth map is obtained through neighborhood point matching, and therefore each point needs to be refined; thus, the implementation of semi-supervised frameworks becomes more difficult using the PatchMatch method.…”