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

Bad texture after reconstructing colored mesh #89

Closed Kasidit0052 closed 3 years ago

Kasidit0052 commented 3 years ago

Hello,

First, Thank you for creating this repository. Currently, I am trying to reconstruct colored mesh using your Colab. However, there are an unwanted artifacts on the output result. From the image below, figure coat are wrongly color when sigma=20. When reducing sigma, it is possible to improve the mesh. However, the output mesh is too noisy in this situation.

result

This is the 3d reconstruction config:

occ_threshold = 0.2
xmin, xmax = -0.5, 0.5 
ymin, ymax = -0.5, 0.5 
zmin, zmax = -2.04, -0.64 
sigma_threshold = 5. &&  20.     

If possible, could you please elaborate the problem since modifying the occluded threshold doesn't help it either and nerf training result seem to be good also (training_psnr = 28.8 and val_psnr = 28.0). I am happy to provide checkpoint and poses_bounds.py if needed.

Thank you!

kwea123 commented 3 years ago

Hi, I think the mesh you get is probably the best we can get from this algorithm. The bad texture problem is a known limitation, and we don't have solution currently. There are some works dealing with texture, and also mesh quality (by removing the dependency on sigma). I might try to implement some of these features in the future, but there is no plan currently.