Why is the max_angle in superpoint+lightglue_homography.yaml set to 45 degrees?

[ChangHsinChih]

When I use two images with large rotation angles, I notice that the matching performance is poor. I'm wondering if increasing the max_angle would improve the performance?

[Phil26AT]

Hey @lgdtd3

Thank you for opening this issue. Indeed in-plane rotations are a limitation of the work, and this is partially inherited from the descriptors. Because we do not train the descriptors, and because most images are roughly upright, we trained the model with 45 degree augmentations (note that this results in up to 90 degree relative rotation for an image pair). Also, most of the images in MegaDepth are upright, too, so you also have to add rotations there.

You could try to improve the rotation robustness of LightGlue by increasing this angle, but I doubt it would solve the issue (it could work with SIFT). Alternatively, you could also manually rotate one of the images by 90/180/270 degrees and select the pair with most matches.

[blackjacket996]

Yeah，using a trick like rotating the image before would bring a better result. I've tried it