NVlabs / neuralangelo

Official implementation of "Neuralangelo: High-Fidelity Neural Surface Reconstruction" (CVPR 2023)
https://research.nvidia.com/labs/dir/neuralangelo/
Other
4.31k stars 387 forks source link

SCENE_TYPE=outdoor make loss=NAN and train stop #110

Open Dragonkingpan opened 11 months ago

Dragonkingpan commented 11 months ago

I have an outdoor scene where the camera is not rotating around an object, but rather walking a long distance around a building. I set the parameter SCENE_ TYPE=exterior. As a result, when I took the first step of training, there was a loss=NAN error, and the training was terminated. I will call this parameter back to the object, and this error will no longer occur. I would like to ask what the meaning of this parameter is, and will using the object in my outdoor scene cause any other errors?

mli0603 commented 11 months ago

Hi @Dragonkingpan

The scene_type has two use cases:

Can you check which of the above leads to your problem? My guess is that the bounding box is the issue. For outdoor scenes, the camera trajectory can be arbitrary, so there is very little prior info that we can use in comparison to an object-centric scene. The script tries its best to bound the scene, however, sometimes this is not sufficient.