chenhsuanlin / bundle-adjusting-NeRF

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

why it tell me the warning (SVD did not converge) when i train the model used my data #64

Closed YS-IMTech closed 1 year ago

YS-IMTech commented 1 year ago

why it tell me the warning (SVD did not converge) when i train the model used my data

Below for details: image

chenhsuanlin commented 1 year ago

It is because Procrustes analysis failed (see this line). It could be because there is only one point being aligned, all the points are completely overlapping (resulting in a division-by-zero), or other causes. Hope this helps!

jonathanhyunmoon commented 1 year ago

Is this warning a problem for training? It persists throughout the training. Here are my loss and psnr curves.

Screenshot from 2023-01-18 09-25-28

Screenshot from 2023-01-18 09-24-31

chenhsuanlin commented 1 year ago

Hi @jonathanhyunmoon, if you are running train-from-scratch experiments, it will raise the warning at the very beginning as the initial poses are completely overlapping (all identity), thus resulting in a division-by-zero. It shouldn't happen again once the training starts. If that's the case for you, then it should be fine. For @YS-IMTech's case, there seem to be other issues.

chenhsuanlin commented 1 year ago

Closing due to inactivity, please feel free to reopen if necessary.