Keypoint Scale for HPatches?
xushangnjlh opened this issue · 4 comments
Thank you for sharing this wonderful work!
I am testing your results on HPatches, however I am not sure which keypoint scale (sampling window) you utilized when testing BEBLID descriptor. I noticed that you have mentioned a recommended scale for different keypoint (so I set 6.75 for hpatches). But the resulting matching mAP is only 0.13.
So what are the scale settings for you when evaluating on hpatches? Thanks in advance!
Hi,
I assume that you are using the OpenCV Python version and that you are using pre-cropped patches. Is this right?
If so, for the 32x32 patches, the keypoint should be centered in the point (16, 16) and the scale=1.0. This is what we have tested in the paper. If you use patches of size 64x64 then the center should be (32, 32) and the scale=2.0. I hope this is helpful for you. If you are not able to reproduce the results and can send you a small example, but I'm not sure when I'll have time to do it.
It would be great if you can send me the example (xushangnjlh@gmail.com). As for HPatches, default patch size is 65x65 and when I set keypoint scale to 2.0, the results get even worse and I cannot figure out why ...
Thanks and take your time~
Okey, I'll try to do it this weekend! I will also publish it here because I think is beneficial for the community.
Best,
Iago.
Hi @xushangnjlh ,
Sorry for late reply. My time is scarce and this type of thing takes time. Here you can find a python notebook where I leave you the instructions to reproduce the paper results.
The figure is exactly the same as the one in the paper and BEBLID obtain the same results.
I hope you find this useful :)
Best,
Iago.