search

hustvl / 4DGaussians

[CVPR 2024] 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering
https://guanjunwu.github.io/4dgs/
Apache License 2.0
2.25k stars 187 forks source link

The performance of the model #114

Open Maxwell-Zhao opened 8 months ago

Maxwell-Zhao commented 8 months ago

I followed the example code to train hypernerf scenes such as virg/broom, but the training results are not as expected. Why is that?

13999_0 13499_0

Maxwell-Zhao commented 8 months ago

QQ图片20240328153734 The model doesn't converge.

Maxwell-Zhao commented 8 months ago

I have tried multiple environments, including py37+torch1.31+cu11.6 and py38+torch2.01+cu11.8. However, the final results are all not satisfactory.

guanjunwu commented 8 months ago

I have tried multiple environments, including py37+torch1.31+cu11.6 and py38+torch2.01+cu11.8. However, the final results are all not satisfactory.

It looks very strange. Can you show me the point clouds generated by colmap? It may look like: image

which is previewed by visual studio code.

Maxwell-Zhao commented 8 months ago

I have tried multiple environments, including py37+torch1.31+cu11.6 and py38+torch2.01+cu11.8. However, the final results are all not satisfactory.

It looks very strange. Can you show me the point clouds generated by colmap? It may look like: image

which is previewed by visual studio code.

This is the point clouds generated by colmap. QQ图片20240329092121 This is the point clouds generated by the model. QQ图片20240329092218

guanjunwu commented 8 months ago

It seems that the original 3D Gaussians do not fit well. In my test, after 3000 coarse iterations, the training PSNR will be up to 20. Have you tested in any other scenes?