About feature extraction accuracy

[Shaohanm]

Hi,

First of all, thanks for this excellent work.

When I use Lightglue for matching, I found the input image resized to 1024 in the default config and want to know the reason for doing this. Because I found that image size changes may cause feature point positions to offset, and I tried sparse reconstruction based on matching results, which may increase mean reprojection errors.

Thank you

[Phil26AT]

Hey @Shaohanm

We think that image resizing is actually bound to the extractor since it affects the localization accuracy of keypoints, and thus we added it to the extract function. The default value 1024px is based on heuristics and experience, and provides a nice trade-off between accuracy and extraction speed. This way, we also guarantee that you can run the default code on very large images without running into memory issues. For higher accuracy, we suggest increasing it to 1600px. You can also disable it:

extractor = SuperPoint(max_num_keypoints=2048).eval().cuda()  # load the extractor
image0 = load_image('path/to/image_0.jpg').cuda()
# extract local features
feats0 = extractor.extract(image0, resize=None)

[Shaohanm]

Thanks for your quick reply, I'm sure your suggestion will work.

I also want to ask another question, how to ensure the accuracy of feature point extraction when the image quality is not good (for example, some images may have noise or ghosting), and I'm looking forward to your suggestions.

Thank you