Stereo matching based on MST and related improvements (A non-local cost Aggregation Method for Stereo Matching)

With a very tangled mood to summarize this paper, this is mainly because the author, although the source code, but I did not carefully go into his code, just put his algorithm added to his project. Hope that there will be time to the MST this structure of their own programming implementation!!

The thesis is based on the non-local cost clustering (non-local costs aggregation) stereo matching, from the topic of this paper is not a local algorithm, but also is not the traditional sense of the global algorithm. This should start from the local stereo matching algorithm based on window structure, such as:

We seek the cost of the left and right two images at the lower point of Parallax D, we are actually finding the average of the cost of all points in Windows with the center radius r. We call this process cost aggregation, and it is clear that a point outside of Windows does not affect costs at that point, which can be very prone to mis-matching in areas with lower texture characteristics. The work of this paper focuses on this cost clustering, the goal is that all the points in the image to the point of transmission of a support, distance from the point, the color difference is very large point of transmission of small support. What should we do if we have a goal? This paper presents the MST structure. According to the MST structure we know that when the image is viewed as a graph of a four Unicom region, the weights of the edges formed by the two points of the image are defined as the difference of the gray values of these two points, and the MST structure generated under this definition coincides with our expectations. This approach is equivalent to adding global properties to local algorithms, so it is called a nonlocal algorithm.

Even more rare is that the algorithm according to the author pushed down the formula only need to the MST traverse two times to get all the points of the cost of clustering, the algorithm complexity is obviously reduced.

The paper is mainly divided into four steps, here is the main cost of clustering process:

After we generate an MST:

According to the formula:

The cost clustering values of all nodes are obtained from the leaf node to the vertex and the vertex to the leaf node two times respectively.

where S (p,q) is defined as a similarity of two points, D (P,Q) is defined as a distance of two points (the smallest path directly to the MST two points), which is the cluster value. The definitions are as follows:

Finally, the disparity map without further optimization can be obtained according to the clustering results.

Finally, based on this paper, the price-change algorithm "Segment-tree based cost Aggregation for Stereo Matching" also obtained a good effect. First, the reference image is based on the segmentation of the graph, and then the resulting sub-tree aggregation using greedy algorithm to converge into a tree, according to the principle, this is also an MST. But in the concrete implementation is encountered some problems:

1, when the image is segmented we use the compression path and look up the set, but each node of the tree will point to the same vertex. 2, sub-tree merging also encountered problems, the paper seems to be directly to the tree vertex merging. These two problems can not be solved to the next step of the work, and then continue to study it!! Finally, attach the article Code link: http://www.cs.cityu.edu.hk/~qiyang/Most of the papers in his homepage are provided with code.

MST-based stereo matching and related improvements (A non-local cost Aggregation Method for Stereo Matching)