False Positive Keypoints on uniform Images
DavideCatto opened this issue · 3 comments
Hello to everyone and ty for your work.
I'm extracting keypoints on uniform color images and these are the results. I have also tried to apply some gaussian blur but nothing change. How resolve this problem?
Input Image
Output Keypoints draw on Input Image
Hi @DavideCatto, thank you for your interest in LightGlue-ONNX.
I don't think the extractors were trained to detect keypoints/features on uniform colour images (though in the first place, how can one obtain meaningful representations from something uniform?), so the result is the random keypoints you're getting.
Hi @fabio-sim, yeah you're right, my question was about how to avoid to have results when you pass an uniform color image to the network. I think it was trained only on images with keypoints so when I pass an image outside the distribution of train images to the model, random keypoints are extracted. Putting a threshold on disk's keypoints not seem to resolve the problem. I have also check the orginal implementation but same results.
I think the only solution is to train disk model with also images with no keypoints.
I agree. You would have to train the extractor to explicitly ignore uniform images.