OpenCV SIFT+FLANN multiple matches for single keypoint

Question

I'm trying to match keypoints using open cv. Specifically, I'm using the "sift" detector and "flann" matcher. My code is based on cv2's documentation:

detector = cv2.SIFT_create()
matcher = cv2.FlannBasedMatcher(indexParams=dict(algorithm=0, trees=5), searchParams=dict(checks=50))
kps1, desc1 = detector.detectAndCompute(img1, None)
kps2, desc2 = detector.detectAndCompute(img2, None)
all_matches = matcher.knnMatch(desc1, desc2, 2)

ratio = 0.7
good_matches = []
for m, n in all_matches:
  if m.distance <= ratio * n.distance:
    good_matches.append(m)

I noticed than even within the good_matches list, I have some keypoints that have more than a single match:

extra_matches = dict()
for match in good_matches:
  t_idx = match.trainIdx
  reps = [mch for mch in good_matches if mch.trainIdx == t_idx]
  if len(reps) > 1 and t_idx not in extra_matches.dict():
    extra_matches[t_idx] = reps

print(len(extra_matches))  # not 0

I find this weird because I thought that knnMatch already yields the 2 best matches. Why would I have more than a single match per keypoint after ratio-pruning the matches?

Answer 1

Sure enough, five minutes after posting I found the answer:

FLANN does not do cross-checks, which means I will have repeats of the 2nd keypoints but no repeats for the 1st keypoints (verified in my code as well). The best practice if you need cross-check with FLANN is to implement your own cross-check or use FLANN to get a subset of descriptors and then use BFMatcher's cross-check option.

Here are some other sources of information: [1], [2]