search

zju3dv / LoFTR

Code for "LoFTR: Detector-Free Local Feature Matching with Transformers", CVPR 2021, T-PAMI 2022
https://zju3dv.github.io/loftr/
Apache License 2.0
2.34k stars 365 forks source link

Details of reproducing InLoc localisation #150

Closed vdvchen closed 2 years ago

vdvchen commented 2 years ago

Hi, thanks for your great works.

I'm trying to reproduce your results on InLoc localisation. As is suggested in paper, LoFTR is combined with HLoc to perform visual localisation. I integrate LoFTR into the inloc pipeline and find some performance gap compared with the reported in paper.
Would you mind sharing some experiment details on InLoc reproduction, that should be very helpful!

I understand some settings are elaborated on the leaderboard(https://www.visuallocalization.net/details/19843/), like using pairs-query-netvlad40-temporal.txt as retrieval list and using outdoor model(please correct me if I'm wrong). But some other details like resize and threshold should also matter.

Thanks!

zehongs commented 2 years ago

Hi, I think the major difference is that we used the finetuned version of the outdoor-OT model, the skh-iter is set as 3 and the confidence threshold is set as 0.2. The augmentation is as follows. https://github.com/zju3dv/LoFTR/blob/2122156015b61fbb650e28b58a958e4d632b1058/src/utils/augment.py#L4-L19

vdvchen commented 2 years ago

Thanks for your help!