Differernt SIFT descriptor as input to matcher

[gunjanmimo]

I am exploring the use of COLMAP's descriptor with SIFT_ROOT normalization and VLFEAT's SIFT descriptor with SIFT_ROOT normalization. My initial tests show that both are performing correctly. However, I am opening this issue to understand whether SIFT descriptors from different sources, such as COLMAP or VLFEAT, will function as expected with LIGHTGLUE's SIFT implementation.

Specifically, I would like to know:

• Will using the COLMAP or VLFEAT SIFT descriptors work seamlessly with LIGHTGLUE as it does with LIGHTGLUE's native SIFT?
• Are there any known compatibility issues or performance differences when using these third-party SIFT descriptors with LIGHTGLUE?

I appreciate any insights or experiences regarding this matter.

Thank you!

[Phil26AT]

Hey @gunjanmimo, COLMAP actually uses VLFEAT (see here), so both of them should work. LightGlue also works with OpenCV SIFT.

LightGlue was trained with COLMAP SIFT-GPU, and thus this usually performs best.

I tested it with kornia SIFT, but unfortunately this does not work so far. If you happen to make them compatible, feel free to open a PR, I think many people would love to see that!