Training immediately converges to white.

kwea123 / nerf_pl

NeRF (Neural Radiance Fields) and NeRF in the Wild using pytorch-lightning

https://www.youtube.com/playlist?list=PLDV2CyUo4q-K02pNEyDr7DYpTQuka3mbV

MIT License

2.74k stars 483 forks source link

Training immediately converges to white. #39

Closed sungam94 closed 4 years ago

sungam94 commented 4 years ago

Has anyone experienced a data set immediately (few hundred iterations) converging to white? It's a 360 of a figurine like kwea123's featured in the README of this repo; with 50 photographs taken in front of a white background. COLMAP has done a successful solve with all cams. Does anyone have an idea how to debug?

kwea123 commented 4 years ago

How do the images look like? Does the white background occupy most of the area?

sungam94 commented 4 years ago

Hi kwea123, thanks for sharing your work and the fast reply. Here is a screenshot: archer

kwea123 commented 4 years ago

Are you sure COLMAP has successfully resolved the poses? I've heard that too clean white background has a negative impact on the pose solution... you can check this using gui for example. If the poses are indeed correct, then maybe try to increase the batch size (so that more samples from the object are drawn) or try other optimizers such as radam or ranger. A little more complicated solution can be found here