I have a set of templates images against which I need to compare a test image and find the best match. Given that we have a SIFT descriptor, I select the best feature match and all feature matches that lie within 3*distance of the best match are considered good matches. Then I add up the distance of all the good matches. I don't know if this is a good approach because I think I should also take into the account the number of good matches, not just the average of the distances between the good matches. I am new to template matching, so I would appreciate your inputs.
0
votes
1 Answers
1
votes
In these test images, is the template that you are looking for always in the same perspective (undistorted)? If so, I would recommend a more accurate technique than using feature point matching. OpenCV offers a function called matchTemplate() and there is even a gpu implementation. Your measure can be based on the pixel averaged result of that function.
If they are distorted, then using SIFT or SURF might suffice. You should send your point matches through findHomography(), which will use RANSAC to remove outliers. The number of matches that survive this test could be used as a measure to decide if the object is found.
