This is the offical Pytorch implementation of the paper:"

ARTS: Semi-Analytical Regressor using Disentangled Skeletal Representations for Human Mesh Recovery from Videos (ACM MM 2024)

Tao Tang, Hong Liu, Yingxuan You, Ti Wang, Wenhao Li,

arXiv / paper

Preparation

1. Install dependencies. This project is developed on Ubuntu 18.04 with NVIDIA 3090 GPUs. We recommend you to use an Anaconda virtual environment.

# Create a conda environment.
conda create -n arts python=3.8
conda activate arts

# Install PyTorch >= 1.2 according to your GPU driver.
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge

# Install other dependencies.
sh requirements.sh

2. Prepare SMPL layer.

• For the SMPL layer, We used smplpytorch. The repo is already included in ./smplpytorch folder.
• Download basicModel_f_lbs_10_207_0_v1.0.0.pkl, basicModel_m_lbs_10_207_0_v1.0.0.pkl, and basicModel_neutral_lbs_10_207_0_v1.0.0.pkl from SMPL (female & male) and SMPL (neutral) to ./smplpytorch/smplpytorch/native/models.

Implementation

Data Preparation

Rename the ./data_final to the ./data. And the ./data directory structure should follow the below hierarchy. Download all the processed annotation files from OneDrive

${Project}  
|-- data  
|   |-- base_data
|   |   |-- J_regressor_extra.npy
|   |   |-- mesh_downsampling.npz
|   |   |-- smpl_mean_params.npz
|   |   |-- smpl_mean_vertices.npy
|   |   |-- SMPL_NEUTRAL.pkl
|   |   |-- spin_model_checkpoint.pth.tar
|   |-- COCO  
|   |   |-- coco_data  
|   |   |-- __init__.py
|   |   |-- dataset.py
|   |   |-- J_regressor_coco.npy
|   |-- Human36M  
|   |   |-- h36m_data  
|   |   |-- __init__.py
|   |   |-- dataset.py 
|   |   |-- J_regressor_h36m_correct.npy
|   |   |-- noise_stats.py
|   |-- MPII  
|   |   |-- mpii_data  
|   |   |-- __init__.py
|   |   |-- dataset.py
|   |-- MPII3D
|   |   |-- mpii3d_data  
|   |   |-- __init__.py
|   |   |-- dataset.py
|   |-- PW3D 
|   |   |-- pw3d_data
|   |   |-- __init__.py
|   |   |-- dataset.py
|   |-- multiple_datasets.py

Quick Demo

1. Install ViTPose. ARTS uses the off-the-shift 2D pose detectors to detect persons from images. Here we take and install ViTPose.

git clone https://github.com/open-mmlab/mmcv.git
cd mmcv
git checkout v1.3.9
MMCV_WITH_OPS=1 pip install -e .
(or pip install mmcv-full==1.5.0 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html)
cd ..
git clone https://github.com/ViTAE-Transformer/ViTPose.git
cd ViTPose
pip install -v -e .

2. Download the pre-trained ViTPose model vitpose-h-multi-coco.pth from OneDrive. Put it below ./pose_detector folder.
3. Download the pre-trained ARTS model ARTS_Demo_Model and ARTS_PoseEstimation from GoogleDrive. Put it below ./experiment folder.
4. Prepare the input video *.mp4 and put it below ./demo folder.
5. Run. The output is at ./output folder.

# Change 'sample_video' to your video name.
sh ./main/demo.sh 
or
python ./main/run_demo.py --vid_file demo/sample_video.mp4 --gpu 0

Train

Stage 1 : Train the 3D pose estimation.

# Human3.6M
bash train_pose_h36m.sh

# 3DPW
bash train_pose_3dpw.sh

Stage 2: To train the all network for final mesh. Configs of the experiments can be found and edited in ./config folder. Change posenet_path in ./config/train_mesh_*.yml to the path of the pre-trained pose model.

# Human3.6M
bash train_mesh_h36m.sh

# 3DPW & MPII3D
bash train_mesh_3dpw.sh

Test

To test on a pre-trained pose estimation model (Stage 1). Download the pre-trained ARTS model ARTS_Demo_Model(3DPW Dataset) and ARTS_PoseEstimation(3DPW Dataset) from GoogleDrive. Put it below ./experiment folder.

# Human3.6M
bash test_pose_h36m.sh

# 3DPW
bash test_pose_3dpw.sh

To test on a pre-trained mesh model (Stage 2).

# Human3.6M
bash test_mesh_h36m.sh

# 3DPW
bash test_mesh_3dpw.sh

# MPII3D
bash test_mesh_mpii3d.sh

Change the weight_path in the corresponding ./config/test_*.yml to your model path.

Citation

Cite as below if you find this repository is helpful to your project:

@inproceedings{tang2024arts,
  title={ARTS: Semi-Analytical Regressor using Disentangled Skeletal Representations for Human Mesh Recovery from Videos},
  author={Tang, Tao and Liu, Hong and You, Yingxuan and Wang, Ti and Li, Wenhao},
  booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
  pages={1514--1523},
  year={2024}
}

Acknowledgement

This repo is extended from the excellent work PMCE, Pose2Mesh, TCMR. We thank the authors for releasing the codes.