D-PoSE: Depth as an Intermediate Representation for 3D Human Pose and Shape Estimation

Pre-Print

Install

Docker installed and properly configured NVIDIA Container Toolkit installed

• CUDA GPU required D-PoSE model files:
• data/ckpt/paper_arxiv.ckpt
• Google Drive link Register for SMPL and SMPLX body models (READ LICENSE):
• https://smpl.is.tue.mpg.de/
• https://smpl-x.is.tue.mpg.de/

docker build -t dpose .
docker compose -p dpose-0 up -d
docker exec -it dpose-0 bash
pip install git+https://github.com/nikosvasilik/neural_renderer
cd dpose

D-PoSE demo

 python3 demo.py --cfg configs/dpose_conf.yaml

Evaluation

Checkpoint and more instructions coming soon. Default dataset for evauluation is 3DPW. Change dpose_conf.yaml VAL_DS value to change the testing dataset.

 python3 train.py --cfg configs/dpose_conf.yaml --ckpt data/ckpt/paper.ckpt --test

Training

 python3 train.py --cfg configs/dpose_conf.yaml

Qualitative Results

Citation

@article{vasilikopoulos2024d,
  title={D-PoSE: Depth as an Intermediate Representation for 3D Human Pose and Shape Estimation},
  author={Vasilikopoulos, Nikolaos and Drosakis, Drosakis and Argyros, Antonis},
  journal={arXiv preprint arXiv:2410.04889},
  year={2024}
}

References

We benefit from many great resources including but not limited to BEDLAM,SMPL-X,TokenHMR,SMPL, PARE,ReFit ,CLIFF, AGORA, PIXIE, HRNet.