I think there is a bug in subpixel prediction

[CodeRed97]

Thanks for your work and sharing! I think I found a bug while learning your paper and your code: In utils.losses.py, function 'extract_patch_from_points', it seemed the patch extracted does not take the predicted integer coodinate as the center, but take it as the left-top corner, which doesn't make sense in theory. This may cause the prediction accuracy with subpixel poorer than without subpixel. However, this can be easily corrected with some changes in utils.losses.py, function 'extract_patch_from_points' and 'soft_argmax_2d'.(already tested) I hope I didn't make a mistake and this could be helpful. Thanks for your contribution again.

[P66094125]

CodeRed97 1. Can you show your modified code? thank you .2.how do you set the config to train the subpixel superpoint&magicpoint? thank you .

[foxkw]

Thanks for your work and sharing! I think I found a bug while learning your paper and your code: In utils.losses.py, function 'extract_patch_from_points', it seemed the patch extracted does not take the predicted integer coodinate as the center, but take it as the left-top corner, which doesn't make sense in theory. This may cause the prediction accuracy with subpixel poorer than without subpixel. However, this can be easily corrected with some changes in utils.losses.py, function 'extract_patch_from_points' and 'soft_argmax_2d'.(already tested) I hope I didn't make a mistake and this could be helpful. Thanks for your contribution again.

res_ext = torch.cat(((1-x_res)*(1-y_res), (1-x_res)*y_res, x_res*(1-y_res), x_res*y_res), dim=0) Is the design of this res unreasonable? The corresponding distance between the approximate points should be smaller