search

georg-bn / rotation-steerers

A steerer for D-dimensional keypoint descriptions is a DxD matrix that transforms the descriptions as if they were computed from a rotated image.
MIT License
71 stars 8 forks source link

Self-supervised training #6

Closed guipotje closed 5 months ago

guipotje commented 5 months ago

Hi, thank you for open-sourcing the code, really liked the idea of the paper.

Just wondering if you tried to optimize the projection matrix based on proximity in the descriptor space instead of supervising via softmax matching (for example L2 or cosine similarity). This would allow easy training setup with just a dataset of random images. If you didn't, do you think this approach would work well?

Best,

Guilherme.

georg-bn commented 5 months ago

Yes, I have tried that and it usually works quite well! I think L1 distance in the descriptor space worked better than L2 when I was experimenting. In the end we used the matching-based training in the paper just for consistency with setting B and C .

guipotje commented 5 months ago

Great to know it works well! Thank you!