1, in the binocular matching, the feature points are matched by gray-level correlation, so there is inevitably a false matching point pair. (I understand that the image is the same point in both images by the same point in the gray values of the feature points in the left and right images, and the pixels in the whole image may have a lot of points, but in practice they are not the same in the camera, so there will be a false match)
2, at this time, in order to reduce this mismatch, we have to give some constraints, that is, under certain conditions, the pixel value can be considered to be a matching knot. These conditions are called constraints, so there is the argument that there is a polar line constraint, also known as the polar geometry.
3. Matching classification:
Local constraint-based methods: There are region matching (mainly based on window), feature matching (based on feature points, such as sift), phase matching is also called gradient method (mainly by filtering).
The method based on global constraint includes dynamic programming algorithm, graph cutting algorithm, artificial intelligence algorithm, cooperative algorithm, confidence propagation algorithm and nonlinear diffusion algorithm.
There is no place for everyone to discuss with me.