Avatars Grow Legs: Generating Smooth Human Motion from Sparse Tracking Inputs with Diffusion Model Y. Du, R. Kips, A. Pumarola, S. Starke, A. Thabet, A. Sanakoyeu CVPR 2023
All our experiments are done on a single V-100 16G GPU.
conda env create -f environment.yml
conda activate agrol
The code was tested on Python 3.9 and PyTorch 1.12.1.
Download the human_body_prior lib and body_visualizer lib and put them in this repo. The repo should look like
agrol
├── body_visualizer
├──── mesh/
├──── tools/
├──── ...
├── human_body_prior/
├──── body_model/
├──── data/
├──── ...
├── dataset/
├── prepare_data/
└── ...
Please download the AMASS dataset from here(SMPL+H G).
python prepare_data.py --support_dir /path/to/your/smplh/dmpls --save_dir ./dataset/AMASS/ --root_dir /path/to/your/amass/dataset
The generated dataset should look like this
./dataset/AMASS/
├── BioMotionLab_NTroje
├──── train/
├──── test/
├── CMU/
├──── train/
├──── test/
└── MPI_HDM05/
├──── train/
└──── test/
You can either download our pre-trained models or use your pre-trained model. To download our pre-trained models:
sh prepare_data/download_model.sh
To evaluate the model:
# Diffusion model
python test.py --model_path /path/to/your/model --timestep_respacing ddim5 --support_dir /path/to/your/smpls/dmpls --dataset_path ./dataset/AMASS/
# MLP
python test.py --model_path /path/to/your/model --support_dir /path/to/your/smpls/dmpls --dataset_path ./dataset/AMASS/
To train the AGRoL diffusion-model:
python train.py --save_dir /path/to/save/your/model --dataset amass --weight_decay 1e-4 --batch_size 256 --lr 3e-4 --latent_dim 512 --save_interval 1 --log_interval 1 --device 0 --input_motion_length 196 --diffusion_steps 1000 --num_workers 8 --motion_nfeat 132 --arch diffusion_DiffMLP --layers 12 --sparse_dim 54 --train_dataset_repeat_times 1000 --lr_anneal_steps 225000 --overwrite
To train the MLP model:
python train.py --save_dir /path/to/save/your/model --dataset amass --weight_decay 1e-4 --batch_size 256 --lr 3e-4 --latent_dim 512 --save_interval 1 --log_interval 1 --device 0 --input_motion_length 196 --diffusion_steps 1000 --num_workers 8 --motion_nfeat 132 --arch mlp_PureMLP --layers 12 --sparse_dim 54 --train_dataset_repeat_times 1000 --lr_anneal_steps 225000 --overwrite --no_normalization
The pretrained weights for AGRoL can be downloaded from this link: https://github.com/facebookresearch/AGRoL/releases/tag/v0
To download the wights automatically, please run bash prepare_data/download_model.sh
.
To test the pretrained AGRoL diffusion-model:
python test.py --model_path pretrained_weights/diffmlp.pt --timestep_respacing ddim5 --support_dir /path/to/your/smpls/dmpls --dataset_path ./dataset/AMASS/
To visualize the generated motions, add these commands behind:
--vis --output_dir /path/to/save/your/videos
The majority of AGRoL code is licensed under CC-BY-NC, however portions of the project are available under separate license terms:
- Trimesh, AvatarPose, Guided Diffusion, and MDM are licensed under the MIT license;
- Human Body Prior is licensed under a custom license for non-commercial scientific research purposes, available at link;
- Body Visualizer is licensed under a custom license for non-commercial scientific research purposes, available at link.
If you find our work inspiring or use our codebase in your research, please consider giving a star ⭐ and a citation.
@inproceedings{du2023agrol,
author = {Du, Yuming and Kips, Robin and Pumarola, Albert and Starke, Sebastian and Thabet, Ali and Sanakoyeu, Artsiom},
title = {Avatars Grow Legs: Generating Smooth Human Motion from Sparse Tracking Inputs with Diffusion Model},
booktitle = {CVPR},
year = {2023},
}
If you encounter this error during visualization:
ValueError: Cannot use face colors with a smooth mesh
You can fix it by changing the line 88 in your body_visualizer/mesh/mesh_viewer.py
to:
mesh = pyrender.Mesh.from_trimesh(mesh, smooth=False)