search

crockwell / far

[CVPR 2024 - Highlight] FAR: Flexible, Accurate and Robust 6DoF Relative Camera Pose Estimation
https://crockwell.github.io/far/
101 stars 6 forks source link

mp3d_loftr, demo #7

Closed SongMingtao-NUDT closed 2 months ago

SongMingtao-NUDT commented 2 months ago

Hello, when I run the demo in mp3d_loftr, the result I get is different from the "expected prediction" in 'demo.sh'. The result I get is: [[ 0.0201 -0.1913 0.9813 -0.4618] [ 0.1811 0.966 0.1846 -0.1645] [-0.9833 0.174 0.054 0.8716]] But the "expected prediction" in 'demo.sh' is : [ 0.0367 -0.1899 0.9811 -0.4666] [ 0.1938 0.9645 0.1795 -0.1628] [-0.9803 0.1836 0.0722 0.8693] I want to know if these errors are within allowable limits. If not, is my result wrong or the "expected prediction" wrong?

crockwell commented 2 months ago

Hi Songming,

Ahh that is a bit pesky. I'd say that's within reasonable limits. If I had to guess what is going on, perhaps PyTorch or other versioning is slightly different than what I ran for this (I used PyTorch 1 Anaconda env for mp3d)? In any case, you are still closely replicating results. If you're concerned about benchmarking, you could try re-running evaluation. I bet it would be very close. Though let me know if you are concerned.

Best, Chris

SongMingtao-NUDT commented 2 months ago

Hi Songming,

Ahh that is a bit pesky. I'd say that's within reasonable limits. If I had to guess what is going on, perhaps PyTorch or other versioning is slightly different than what I ran for this (I used PyTorch 1 Anaconda env for mp3d)? In any case, you are still closely replicating results. If you're concerned about benchmarking, you could try re-running evaluation. I bet it would be very close. Though let me know if you are concerned.

Best, Chris

Yea, just as you guessed, the result of the re-running evaluation is very close to the data in the paper. Thank you for your reply very much.