search

ingra14m / Deformable-3D-Gaussians

[CVPR 2024] Official implementation of "Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction"
https://ingra14m.github.io/Deformable-Gaussians/
MIT License
779 stars 40 forks source link

Why did I follow the process and get poor results? #64

Open latethousandyear opened 6 days ago

latethousandyear commented 6 days ago

Hello! I did the experiment according to each step of the readme document, and the experimental environment was configured without problems, but the results I got were very bad, and the PSNR was only about 10, what is the reason? poor_result poor_result1

ingra14m commented 2 days ago

Hi, thanks for your interset.

The camera pose plays an important role in Deformable-GS. Did you run on your own dataset or the scene mentioned in the README?

latethousandyear commented 2 days ago

Hi, thanks for your interset.

The camera pose plays an important role in Deformable-GS. Did you run on your own dataset or the scene mentioned in the README?

the scene mentioned in the README

ingra14m commented 3 hours ago

Maybe you used the HyperNeRF dataset? HyperNeRF is a dicey dataset with inaccurate camera pose. The problem with this dataset has been discussed in the paper. Since Deformable-GS can predict accurate and clean motion at a specific time. If the camera pose is not so accurate, it will fail to converge.