XFeat performs significantly worse than SuperPoint on repetitive textures

[DoMoCo]

Hi, thanks for the great work!

I tested XFeat and SuperPoint using vismatch
on repetitive-texture scenes. In my experiments, XFeat performs significantly worse than SuperPoint, especially in feature matching stability.

Superpoint + Lightglue:

XFeat+Lightglue:

My setup:
image size: 512×512
number of keypoints: 2048

The performance gap is quite noticeable. Is this expected behavior for XFeat, or could I be doing something wrong in the pipeline or parameter settings?

Thanks!

[NewboySoCool]

Xfeat + LightGlue 怎么跑的？