Skip to content

HumanGPS builds accurate dense correspondences between human images under arbitrary camera viewpoints and body poses.

License

Notifications You must be signed in to change notification settings

ruofeidu/humangps

 
 

Repository files navigation

HumanGPS: Geodesic PreServing Feature for Dense Human Correspondences

Tensorflow implementation of the paper "HumanGPS: Geodesic PreServing Feature for Dense Human Correspondences", CVPR 2021.

Setup

  • Python 3.6
  • TensorFlow 2.0
  • Tensorflow-Addon
  • gin-config
  • scikit-learn
pip install -r requirements.txt  --user
pip install gdown

Running code

Here we show how to run our code on SMPL intra and inter testing data. You can download the rest of the synthetic SMPL testing data used in the paper here.

1. Download pretrained model

bash download_model.sh

2. Evaluate on intra testing data.

(a) Run

mkdir -p ./test_data/smpl_intra_test/

Download our SMPL intra test data from smpl_intra_data in ./test_data/smpl_intra_test/

To evaluate average epe on intra test dataset.

(b) set JOB_NAME="eval_optical_flow_intra" in ./script/inference_local.sh

(c) Run

bash ./script/inference_local.sh

3. Evaluate on inter testing data.

(a) Run

mkdir -p ./test_data/smpl_inter_test/

Download our SMPL inter test data from smpl_inter_data in ./test_data/smpl_inter_test/

To evaluate average epe on inter test dataset.

(b) set JOB_NAME="eval_optical_flow_inter" in ./script/inference_local.sh

(c) Run

bash ./script/inference_local.sh

4. Inference on toy examples for visualization

Check out ./inference_demo.ipynb for toy examples.

Citation

If you find this code useful in your research, please cite:

@inproceedings{tan2021humangps,
  title = {{HumanGPS: Geodesic PreServing Feature for Dense Human Correspondence}},
  author = {Tan, Feitong and Tang, Danhang and Dou, Mingsong and Guo, Kaiwen and Pandey, Rohit and Keskin, Cem and Du, Ruofei and Sun, Deqing and Bouaziz, Sofien and Fanello, Sean and Tan, Ping and Zhang, Yinda},
  booktitle = {2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2021},
  publisher = {IEEE},
}

About

HumanGPS builds accurate dense correspondences between human images under arbitrary camera viewpoints and body poses.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 87.6%
  • Python 10.7%
  • Other 1.7%