conda create -n geo4d python=3.8.5
conda activate geo4d
pip install -r requirements.txtInstall Pytorch3D
pip install "git+https://github.com/facebookresearch/pytorch3d.git"Download model checkpoint
mkdir checkpoints
mkdir checkpoints/geo4d
gdown 10SPKkOpou2lKl9bwkgx1d6YocYkmSxQl -O ./checkpoints/geo4d/ # fine-tuned vae model
gdown 11K0ubqytun-SA5RIOgR7ejNIR8B4uois -O ./checkpoints/geo4d/ # whole modelWe provide a demo video for you to try our model. Run the inference script:
bash ./scripts/infer_geo4d.sh ./data/demo/drift-turn.mp4 0Please first refer to the evaluation_script.md to download the evaluation datasets.
Then, run the evaluation script:
bash scripts/eval_geo4d.sh [sintel|bonn|kitti|tum|davis] gpu_id
# e.g. bash scripts/eval.sh sintel 0First, install 4d visualization tool, viser.
pip install -e viserYou could then use the viser to visualize the results:
python viser/visualizer.py --data path_to_results_folder --no_mask
We have used codes from other great research work, including DuST3R, MonST3R, DepthCrafter, DynamiCrafter, RayDiffusion, and MoGe. We sincerely thank the authors for their awesome work!
Here are more recent 3D/4D reconstruction projects from our team:
- Dynamic Point Maps: A Versatile Representation for Dynamic 3D Reconstruction
- Flash3D: Feed-Forward Generalisable 3D Scene Reconstruction from a Single Image
- Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images
- VGGT: Visual Geometry Grounded Transformer
If you find Geo4D useful for your research and applications, please cite us using this BibTex:
@misc{jiang2025geo4d,
title={Geo4D: Leveraging Video Generators for Geometric 4D Scene Reconstruction},
author={Zeren Jiang and Chuanxia Zheng and Iro Laina and Diane Larlus and Andrea Vedaldi},
year={2025},
eprint={2504.07961},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2504.07961},
}