search

cvg / LightGlue

LightGlue: Local Feature Matching at Light Speed (ICCV 2023)
Apache License 2.0
3.32k stars 321 forks source link

Issue with SuperPoint and LightGlue Keypoint Matching #83

Open p890040 opened 11 months ago

p890040 commented 11 months ago

Hello,

I've been utilizing the default SuperPoint for feature extraction combined with LightGlue for keypoint matching in this repository, and the results have been quite satisfactory in my use case.

However, today I replaced the SuperPoint weights with ones sourced from https://github.com/rpautrat/SuperPoint. Upon inspection, the model appears to identify a considerable number of keypoints accurately. Nevertheless, when these keypoints and descriptors are fed into LightGlue, it fails to find any matching keypoints.

Is this outcome expected, or could there be some unforeseen issue I've overlooked? I'd appreciate any insights or recommendations regarding this situation.

Thank you.

sarlinpe commented 11 months ago

LightGlue models trained for one kind of descriptors are not compatible with other descriptors. These are 2 different implementations of SuperPoint with very different weights. We will release one model of LightGlue trained for this SuperPoint but this will have to wait longer.

p890040 commented 11 months ago

LightGlue models trained for one kind of descriptors are not compatible with other descriptors. These are 2 different implementations of SuperPoint with very different weights. We will release one model of LightGlue trained for this SuperPoint but this will have to wait longer.

Thanks for the explanation, I'm looking forward to your masterpiece.

nitinagarwal commented 10 months ago

@sarlinpe it will be interesting to see the comparison of the official SPSG and SP(https://github.com/rpautrat/SuperPoint.) + LightGlue. Thanks

avdhoeke commented 5 months ago

Hey, any updates on LightGlue trained for https://github.com/rpautrat/SuperPoint?