This is the implementation of prior-driven initialization described in "3.2.1 Single frame initialization" in "GFlow: Recovering 4D World from Monocular Video".
This initialization method, which takes an image and its corresponding depth as input, can achieve faster 3DGS reconstruction compared to random initialization.
Method | RGB | Depth | Centers |
---|---|---|---|
Prior-init (500steps) | training_rgb_prior.mp4 |
training_depth_prior.mp4 |
training_center_prior.mp4 |
Random-init (5000 steps) | training_rgb_random.mp4 |
training_depth_random.mp4 |
training_center_random.mp4 |
- Install following msplat's instructions.
- run
python tutorials/run.py
and check the results in thelogs
folder.
We thank the msplat team for re-implementing 3DGS in a more developer-friendly way. We also salute the original 3DGS team for their seminal work.
If you find this initialization method helpful, please consider to cite:
@article{wang2024gflow,
title={GFlow: Recovering 4D World from Monocular Video},
author={Wang, Shizun and Yang, Xingyi and Shen, Qiuhong and Jiang, Zhenxiang and Wang, Xinchao},
journal={arXiv preprint arXiv:2405.18426},
year={2024}
}