Colmar-zlicheng/Color-NeuS

Visualization and Mesh is not same

Closed this issue · 4 comments

Talegqz commented

Thanks for your excellent work!

I found some config bug and fixed it, then I ran it with my own data. I found that the visualization is very nice, but the mesh quality is not reasonable.

Could you please explain it or help me improve it?

Thanks for your help.

Snipaste_2023-08-24_11-33-33 Snipaste_2023-08-24_11-33-43
Colmar-zlicheng commented

Hi, that is because the resolution for reconstruction, the default value is 512. You can set it higher when doing evaluation. I think you can set it to 1024 to extract a more refined mesh, while it will take more time for inference.

python evaluation.py -rr 1024 --(other args)
Talegqz commented

Thanks for your help！

xsl316 commented

Thanks for your excellent work!
How to train your own dataset, including how to preprocess your own data！
Thanks for your help.

Colmar-zlicheng commented

You can use the command below to train the IHO-Video dataset, please refer to readme for details.

python train.py -g 0 --cfg config/Color_NeuS_iho.yml -obj ghost_bear --exp_id Color_NeuS_iho_ghost_bear

For data preprocessing, we use colmap to obtain camera poses and TrackAnything to obtain image masks.