Hi @ducha-aiki, thank you for this excellent contribution! This model seems to be a straight improvement over SIFT+LightGlue, which is great.

I think we can simplify the DogHardNet class a bit. Inherit from SIFT, and put these few lines into the forward method (after p=self.extract_singe_image(img)). extract_single_image can then be reused from SIFT, which makes code maintenance easier.

If this is done we can merge, and I can attach your weights to the release.

@Phil26AT Thanks. I can do that, however, there will be a slight inefficiency in the code, as the SIFT descriptors computation will be wasted.

@Phil26AT Done. I have also added missing import into __init__.py and to example in Readme.

Looks great!

@Phil26AT Thanks. I can do that, however, there will be a slight inefficiency in the code, as the SIFT descriptors computation will be wasted.

Yes correct, but this would only apply to opencv SIFT extraction (pycolmap does not allow to turn description off). Did you ever benchmark this? If the difference is big we can maybe pass a flag to run_opencv_sift.

@Phil26AT that is 10ms, which we probably can ignore - the HardNet extraction is dominating the time anyway (esp. on CPU).

Okay yeah I guess we can ignore it for now. I'll merge this PR then, thank you for the contribution!

Great, thank you! Don't forget to name the weights under your convention.

In the assets, dog_hardnet_lightlue should be named doghardnet_lightlue because the dl fails

Fixed, thank you for the note.

@Phil26AT Sorry to bother you again, but it is still wrong at release page:
doghardnet_lightlue.pth instead of doghardnet_lightglue.pth

@ducha-aiki I am sorry for the typo, it is fixed now.

Unreasonable high GPU usage when using DogHardNet . GPU mem leak 160MB to 13 GB .How can i fix it ?