CODEWITHSUNDEEP
pythonopencvcomputer-vision
Hasan Aga
Hasan Aga

Reputation: 774

Matching ORB descriptors, cant find good matches

I am trying to match two images using ORB, using brute force matcher results in random matches, applying Low's ratio condition dose not find any correct match!

I suspect the problem is in if m.distance < 0.7: I read that m.distance is the distance between two discriptors returned from bf matcher, but the values are too large and none meet the condition, how do I remove the outliers then?

What I've tried:

Sorting does not help in finding good matches,and SIFT is patented, I want ORB.

my code:

# finding feature points on products
image = productName + ".jpg"
path = "./images/products/"+image
img1 = cv2.imread(path,cv2.IMREAD_GRAYSCALE) 
orb = cv2.ORB_create()
kps1, des1 = orb.detectAndCompute(img1, None)
result=cv2.drawKeypoints(img1,kps1, img1,flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS )
plt.figure(num=None, figsize=(20, 12), dpi=80, facecolor='w', edgecolor='k')
plt.imshow(result)

# finding feature points on shelves
image = shelfName+'_raf.jpg'
path = "./images/shelves/"+image
img2 = cv2.imread(path,cv2.IMREAD_GRAYSCALE) 
orb = cv2.ORB_create()
kps2, des2 = orb.detectAndCompute(img2, None)
result=cv2.drawKeypoints(img2,kps2, img2,flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
plt.figure(num=None, figsize=(20, 12), dpi=80, facecolor='w', edgecolor='k')
plt.imshow(result)

bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=False)
matches = bf.match(des1, des2)

# store all the good matches as per Lowe's ratio test.
good = []
for m in matches:
    if m.distance < 0.7:
        good.append(m)
        
match_img = cv2.drawMatches(img1, kps1, img2, kps2, good, None, flags=2)

plt.figure(num=None, figsize=(20, 12), dpi=80, facecolor='w', edgecolor='k')
plt.imshow(match_img, 'gray')

enter image description here

enter image description here

enter image description here

if I remove Low's condition, I get: enter image description here

Upvotes: 0

Views: 1621

Answers (1)

Christoph Rackwitz
Christoph Rackwitz

Reputation: 15354

The code attempts to use Lowe's ratio test (see original SIFT paper).

this requires, for every descriptor, the two closest matches.

the code should read:

matches = bf.knnMatch(desCam, desTrain, k=2) # knnMatch is crucial
good = []
for (m1, m2) in matches: # for every descriptor, take closest two matches
    if m1.distance < 0.7 * m2.distance: # best match has to be this much closer than second best
        good.append(m1)

further, I would highly recommend the flann matcher. it's faster than the brute force matcher.

look at the OpenCV tutorials or the samples directory in OpenCV's source (samples/python/find_obj.py) for code that works.

Upvotes: 2

Related Questions