Issue of reproductibility at inference

[leblond14u]

Hi LightGlue community,

I'm facing a weird problem when using LightGlue.
I'm comparing two images of mine :

However when looking at the result I notice 1) poor results and 2) the pattern of the first image features weirdly match the one found on the second :

To be sure it wasn't only on my images I also tried the sacre_coeur test images and those are the results I'm having :


No matches whatsoever ...

I'm using the maximum accuracy setup :

extractor = SuperPoint(max_num_keypoints=None).eval().to(device)
matcher = LightGlue(features='superpoint', depth_confidence=-1, width_confidence=-1).eval().to(device)

Has anyone already faced the same issues or knows how to solve it ?

Thanks in advance,
Best regards,

Hugo

[leblond14u]

Update:
When using LightGlue with DISK features I get nor problems at all. It seems to be linked directly to faulty SuperPoints ...
Weird :/