search

xuelunshen / gim

GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 Spotlight)
https://xuelunshen.com/gim
MIT License
527 stars 24 forks source link

release superglue model ? #2

Closed zebin-dm closed 3 weeks ago

zebin-dm commented 9 months ago

congraduation, thanks for your work. could you please release the superglue model?

xuelunshen commented 9 months ago

Thank you for your attention. The release of the SuperGlue model is under preparation, as we need to consider license and legal issues. We will complete it as soon as possible.

gujiaqivadin commented 9 months ago

Thank you for your attention. The release of the SuperGlue model is under preparation, as we need to consider license and legal issues. We will complete it as soon as possible.

Congratulations for your paper acception! I am wondering whether the weights of superglue model also needs approval of license checking?

xuelunshen commented 9 months ago

Thank you for your attention. The release of the SuperGlue model is under preparation, as we need to consider license and legal issues. We will complete it as soon as possible.

Congratulations for your paper acception! I am wondering whether the weights of superglue model also needs approval of license checking?

@gujiaqivadin Since MagicLeap was involved in propose SuperGlue, it's a commercial company, so we have to be careful with it.

pablovela5620 commented 8 months ago

Thank you for your attention. The release of the SuperGlue model is under preparation, as we need to consider license and legal issues. We will complete it as soon as possible.

Congratulations for your paper acception! I am wondering whether the weights of superglue model also needs approval of license checking?

@gujiaqivadin Since MagicLeap was involved in propose SuperGlue, it's a commercial company, so we have to be careful with it.

Would love to see the lightglue version instead since both the training and inference code are apache 2!

xuelunshen commented 8 months ago

@zebin-dm @gujiaqivadin @pablovela5620

Thank you for your attention; we have now released gim_lightglue and appreciate your patience.

pablovela5620 commented 8 months ago

This is awesome, thank you. Are the gim_lightglue weights a drop in replacement? So if I wanted to use them with the original lightglue repo or in hloc, could I just replace the original weights and things should work? I'd love to do some testing with them

xuelunshen commented 8 months ago

@pablovela5620

Thank you 😊 It may not be possible to directly replace it; it's necessary to reorganize the ckpt file so that the original lightglue can directly read it. Additionally, I found that training lightglue with 2048 points leads to normal performance of gim_lightglue at 2048 points. However, with more points, such as 4096, it might perform strangely because lightglue has not encountered them during training. Therefore, I suggest using the parameters provided by demo.py when utilizing gim_lightglue.