How does matchFeatures Work?

I am right now trying to troubleshoot this function. I understand that there are three metrics that I can use for this function SSD SAD and normxcorr. I have tried all three to match features between stereo images. The problem is that for some reason instead of matching features that are close to one another it overrides itself and matches features that are completely on the opposite side of the image. These are stereo images so the features that are matched should be at most 20 pixels away from each other. Not hundreds of pixels away on the opposite side of the image. Has anyone come across this problem with this function and does anyone know how to fix it?

Answer 1

[Anand]

matchFeatures implements a generic feature matching algorithm. Feature matching for stereo is a much more constrained problem, where you are only matching features in one image to features in the vicinity of that feature in the other image.

This spatial constraint makes it a very different problem.

If your end goal is to find the depth map, look at this example below:

http://www.mathworks.com/help/vision/ug/stereo-vision.html

Another suggestion is to use the epipolar constraint with RANSAC to eliminate outliers in the feature matching as shown in step 4 below:

http://www.mathworks.com/help/vision/gs/stereo-vision.html#btd13cg