4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video

Arxiv | Project Page

Environment Setup

git clone --recursive https://github.com/luoxue-star/4DEquine.git
cd 4DEquine
conda create -n 4DEquine python=3.10
conda activate 4DEquine
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install -e .[all]
pip install "git+https://github.com/facebookresearch/pytorch3d.git"

Inference

Download the checkpoints from here.

First, run AniMoFormer:

python post_optimization_from_video.py --video_path /path/to/video --checkpoint /path/to/checkpoint --cfg /path/to/config --output_dir /path/to/output_folder --stage1 --stage2

After AniMoFormer finishes, run EquineGS:

python demo_avatar.py --animation_params_path /path/to/refined_results.pt --track_mask_file /path/to/mask_list.pkl --img_path /path/to/image --checkpoint /path/to/checkpoint --out_folder /path/to/output_folder

Evaluation

AniMoFormer

Download the dataset JSON files from here.

Generate AniMoFormer predictions:

python post_optimization_from_video.py --video_path /path/to/sequences_folder --checkpoint /path/to/checkpoint --cfg /path/to/config --output_dir /path/to/output_folder --stage1 --stage2

Then set the dataset paths in amr/configs_hydra/experiment/default_val.yaml and run:

python eval_pose.py --config /path/to/config

EquineGS

python eval_avatar.py --checkpoint /path/to/checkpoint --out_folder /path/to/output_folder --image_dir /path/to/image_sequences_dir --postrefine_dir /path/to/post_optimization_output_dir

Training

Download the training data, AniMer backbone and DINOv3 backbone, then update the dataset paths and backbone paths in amr/configs_hydra/experiment/pose.yaml and amr/configs_hydra/experiment/hrm.yaml, respectively.

Train AniMoFormer

bash training_scripts/animoformer.sh

Train EquineGS

bash training_scripts/equinegs.sh

Citation

If you find this code useful for your research, please consider citing the following paper:

@misc{lyu20264dequinedisentanglingmotionappearance,
      title={4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video}, 
      author={Jin Lyu and Liang An and Pujin Cheng and Yebin Liu and Xiaoying Tang},
      year={2026},
      eprint={2603.10125},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.10125}, 
}

Contact

For questions about this implementation, please contact Jin Lyu.