search

naver / r2d2

Other
461 stars 86 forks source link

Could you tell me your training and feature extracting setting for local feature challenge in visual localization benchmark #14

Closed hongsiyu closed 4 years ago

hongsiyu commented 4 years ago

I trained R2D2 following your default setting in your code and paper. But I didn't achieve the best performance on Aachen dataset in your paper.

humenbergerm commented 4 years ago

model_path: r2d2_WASF_N8_big.pt scale_f: 1.189207115002721 min_scale: 0.3 max_scale: 1 min_size: 128 max_size: 9999 max_keypoints: 40000 reliability_thr: 0.7 repeatability_thr: 0.7

jrevaud commented 4 years ago

There is quite a lot of variance on the Aachen, because it depends on RANSAC which is random. If you submit several times the same thing, you'll see that you obtain different results.

This is not a problem of R2D2, it's due to the way the benchmark is implemented and it's the same for all methods.