I am trying to find matching features in two big images (orthophotos covering hundreds of acres at 5 cm resolution) so I can register them. I find many thousands of keypoints in the two images, and then I find cross-checked matches, sort them and take the best ~30% matches. Geographically, the matches are not very good. Here is the code:
def find_matches(key1, dsc1, key2, dsc2):
#Match features - this uses the BFmatcher class, I use it to take advantage of crossCheck
matcher = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
matches = matcher.match(dsc1, dsc2)
# Sort matches by score
matches.sort(key=lambda x:x.distance, reverse=False)
# Remove not so good matches Is this needed when we do geo sort?
print("cross-check matches = ", len(matches))
numGoodMatches = int(len(matches) * GOOD_MATCH_PERCENT)
matches = matches[:numGoodMatches]
# Draw top matches
imMatches = cv2.drawMatches(refimgGray, key1, imgGray, key2, matches, None, flags=4,)
cv2.imwrite(r"C:\AV GIS\matches.jpg", imMatches)
return matches
When I look at a plot of the distance associated with the sorted matches I see many Hamming distance values of the same value: Graph of match distances - x-axis is sorted match number, y-axis is Hamming distance. It looks like stair steps, with many matches sharing the same Hamming distance.
I am adding more filtering of matches to ensure they are geographically close, but I think I am forgoing many good matches just because they are "invisible" to the matcher.
Here is my question: Can anyone explain how matches are decided when there is just one score to evaluate, and it is not unique? I tried the same thing with a single camera image (orders of magnitude smaller in number of pixels) and even then, the distance scores were not unique. Am I misunderstanding how these matches are found and used?
ADDED - here is a graph of best keypoint matches colored by geographic distances. graph of best keypoint matches colored by geographic distances
