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 186 forks source link

Sketchy /rough results on GT poses #100

Open OiOchai opened 8 months ago

OiOchai commented 8 months ago

Hey,

I'm having a weird issue that the results trained on GT camera poses are a lot worse than colmap-estimated ones. Basically, the results with GT poses are rougher/more sketchy than using the colmap pose. The context here is just for static scene, basically for all the dynamic method mentioned here I just train only single frame.

From GT pose: image From colmap-estimated pose: image

To clarify, I still put the GT poses in colmap and follow here to get the pointcloud, and this GT pose and pointcloud works well (and even better than colmap-estimated ones) with:

  1. NeRF/nerfstudio
  2. Dynamic3DGaussians

It works same bad with:

  1. 3DGS original implementation
  2. 4DGaussians

I suspect that something is wrong in camera conversion because both the original 3DGS and your implementation uses the same camera class https://github.com/hustvl/4DGaussians/blob/master/scene/cameras.py#L17. But honestly my cameras are so simple, just pinhole cameras. I cannot think of anywhere could go wrong. Curious to know your thoughts on this! Thank you very much!

guanjunwu commented 8 months ago

Can you provide the dataset link to me and let me try it?