search

chenhsuanlin / bundle-adjusting-NeRF

BARF: Bundle-Adjusting Neural Radiance Fields 🤮 (ICCV 2021 oral)
MIT License
793 stars 114 forks source link

The Procrustes analysis is not robust to scale #74

Open 07hyx06 opened 1 year ago

07hyx06 commented 1 year ago

Hi, thanks for your great work!

I tried to use the provided procrustes_analysis code to align my own optimized camera pose sequence to the ground truth, but found it is not robust to the scale: the aligned pose for [R, t] and [R, 5xt] is different, and thus the resulting R_error and t_error is different (the error for [R, 5xt] is higher than [R, t]). I found that R.det()<0 is True for [R, 5xt], but is False for [R, t]. I wonder if it is caused by numerical precision. Can you help me to make the procrustes_analysis more robust?

chenhsuanlin commented 1 year ago

Hi @07hyx06, this shouldn't happen if you're only scaling the translation. Can you provide a minimal snippet to reproduce such results?