eric-yyjau/pytorch-superpoint

Rotation

adamas-v opened this issue · 3 comments

adamas-v commented

Nice work, but I visualized the matching of Hpatches and superpoint failed in all rotation case.maybe rotation not include in training augmentation?

Deleted user commented

the matching of Hpatches and SuperPoint failed in all rotation case
Me too. My bad case of the result was only two lines in point matching of Hpatches.
I don't know why. Probably, Some of the steps may be lacking in these codes.

adamas-v commented

the matching of Hpatches and SuperPoint failed in all rotation case
Me too. My bad case of the result was only two lines in point matching of Hpatches.
I don't know why. Probably, Some of the steps may be lacking in these codes.

that's beacuse the pre-train model was trained with little sample of rotation, so it is leak of ratotion attack .

Deleted user commented

that's beacuse the pre-train model was trained with little sample of rotation, so it is leak of ratotion attack .

Oh, I see. To deal with problem, it will be better to increase the rotation samples to train the pre-trained model.
These codes didn't mention the step which is training MagicPoint with MS-COCO after Homographic Adaptation. These things may connect to this problem.

  • 👍1