Official PyTroch implementation of HPRNet.

HPRNet: Hierarchical Point Regression for Whole-Body Human Pose Estimation,
Nermin Samet, Emre Akbas,

• Published at IMAVIS. (arXiv pre-print)*

• HPRNet is a bottom-up, one-stage and hierarchical keypoint regression method for whole-body pose estimation.
• HPRNet has the best performance among bottom-up methods for all the whole-body parts.
• HPRNet achieves SOTA performance for the face (76.0 AP) and hand (51.2 AP) keypoint estimation.
• Unlike two-stage methods, HPRNet predicts whole-body pose in a constant time independent of the number of people in an image.

• Results are presented without and with test time flip augmentation respectively.
• All models are trained on COCO-WholeBody train2017 and evaluated on val2017.
• The models can be downloaded directly from Google drive.

0. [Optional but recommended] create a new conda environment.

   conda create --name HPRNet python=3.7
   

   And activate the environment.

   conda activate HPRNet
   

1. Clone the repo:

   HPRNet_ROOT=/path/to/clone/HPRNet
   git clone https://github.com/nerminsamet/HPRNet $HPRNet_ROOT
   

2. Install PyTorch 1.4.0:

   conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
   

3. Install the requirements:

   pip install -r requirements.txt
   

4. First clone the DCNv2 repository into $HPRNet_ROOT/src/lib/models/networks. Then, compile DCNv2 (Deformable Convolutional Networks):

   cd $HPRNet_ROOT/src/lib/models/networks/DCNv2
   ./make.sh
   

• Download the images (2017 Train, 2017 Val) from coco website.

• Download train and val annotation files.

  ${COCO_PATH}
  |-- annotations
      |-- coco_wholebody_train_v1.0.json
      |-- coco_wholebody_val_v1.0.json
  |-- images
      |-- train2017
      |-- val2017 
  

• You could find all the evaluation and training scripts in the experiments folder.
• For evaluation, please download the pretrained models you want to evaluate and put them in HPRNet_ROOT/models/.
• In the case that you don't have 4 GPUs, you can follow the linear learning rate rule to adjust the learning rate.
• If the training is terminated before finishing, you can use the same command with --resume to resume training.

The numerical calculations reported in this paper were fully performed at TUBITAK ULAKBIM, High Performance and Grid Computing Center (TRUBA resources).

HPRNet is released under the MIT License (refer to the LICENSE file for details).

If you find HPRNet useful for your research, please cite our paper as follows:

N. Samet, E. Akbas, "HPRNet: Hierarchical Point Regression for Whole-Body Human Pose Estimation", arXiv, 2021.

BibTeX entry:

@misc{hprnet,
      title={HPRNet: Hierarchical Point Regression for Whole-Body Human Pose Estimation}, 
      author={Nermin Samet and Emre Akbas},
      year={2021}, 
}