Hey there! I see exactly where you're going wrong here—let's break this down and fix your ORB matching problem.

The Core Problem: Misunderstanding ORB's Distance Metric & Lowe's Ratio Test #

First, let's clarify two key points that tripped you up:

• ORB uses Hamming distance (for binary descriptors) instead of the L2 distance SIFT relies on. Hamming distance counts the number of differing bits between two binary vectors, so its value typically ranges from 0 to 64 (since ORB descriptors are 64-bit by default). Comparing this directly to 0.7 makes no sense—you're essentially asking for matches with less than 1 differing bit, which is almost impossible to find.
• You used bf.match() (which returns only the single best match per keypoint) instead of bf.knnMatch()—Lowe's ratio test requires comparing the top two matches for each keypoint to filter out ambiguous, low-quality matches.

Step-by-Step Fixes #

Let's adjust your code and workflow properly:

1. Use k-NN Matching for Valid Lowe's Ratio Testing

Replace your matching code block with this corrected version:

bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=False)
# Get top 2 matches for each keypoint (required for Lowe's ratio test)
matches = bf.knnMatch(des1, des2, k=2)

good = []
# Apply Lowe's ratio test + absolute Hamming distance threshold
for m, n in matches:
    # Ratio test: best match distance / second-best match distance < threshold
    # Add an absolute distance cap to filter out very poor matches
    if m.distance < 0.7 * n.distance and m.distance < 30:
        good.append(m)

2. Tweak ORB Parameters for Better Feature Detection

ORB's default settings might not generate enough high-quality features for your product-shelf matching. Try tuning these parameters to get more robust keypoints:

# Create ORB with increased feature count and optimized settings
orb = cv2.ORB_create(
    nfeatures=2000,  # Default is 500—more features mean more potential matches
    scaleFactor=1.2,
    nlevels=8
)

3. Optional: Try Cross-Checked Matching for More Reliable (Fewer) Matches

If you want tighter, more trustworthy matches (even if fewer in number), enable crossCheck=True in the BFMatcher. This ensures a match is only kept if keypoint A in image 1 matches keypoint B in image 2, and vice versa:

bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
# Cross-check works with match(), not knnMatch()
matches = bf.match(des1, des2)
# Sort matches by distance and keep top N (e.g., top 100)
matches = sorted(matches, key=lambda x: x.distance)[:100]

Why Your Previous Attempts Didn't Work #

• Sorting matches by distance alone doesn't filter out ambiguous matches—you still end up with keypoints that match multiple unrelated spots in the second image.
• Using 0.7 as an absolute distance threshold for Hamming distance is way too strict—most valid ORB matches will have Hamming distances between 10-40, depending on how similar your product and shelf images are.

Final Tips #

• Experiment with the absolute distance threshold (the 30 in the code) based on your specific images: raise it if you're getting too few matches, lower it if clutter still remains.
• Preprocess your images (e.g., apply histogram equalization with cv2.equalizeHist()) to improve feature detection consistency across product and shelf shots.

内容的提问来源于stack exchange，提问作者Hasan Aga