Pose Partition Networks for Multi-Person Pose Estimation

This repository contains the code and pretrained models of

Pose Partition Networks for Multi-Person Pose Estimation [PDF]
Xuecheng Nie, Jiashi Feng, Junliang Xing, and Shuicheng Yan
European Conference on Computer Vision (ECCV), 2018

Prerequisites

• Python 3.5
• Pytorch 0.2.0
• OpenCV 3.0 or higher

Installation

1. Install Pytorch: Please follow the official instruction on installation of Pytorch.
2. Clone the repository

   git clone --recursive https://github.com/NieXC/pytorch-ppn.git
   

3. Download MPII Multi-Person Human Pose dataset and create a symbolic link to the images directory

   ln -s PATH_TO_MPII_IMAGES_DIR dataset/mpi/images
   

Usage

Training

Run the following command to train PPN from scratch with 8-stack of Hourglass network as backbone:

sh run_train.sh

or

CUDA_VISIBLE_DEVICES=0,1 python main.py

A simple way to record the training log by adding the following command:

2>&1 | tee exps/logs/ppn.log

A script to supervise validation accuracy during the training process:

python utils/plot_map_curve.py

Some configurable parameters in training phase:

• -b mini-batch size
• --lr initial learning rate
• --epochs total number of epochs for training
• --snapshot-fname-prefix prefix of file name for snapshot, e.g. if set '--snapshot-fname-prefix exps/snapshots/ppn', then 'ppn.pth.tar' (latest model) and 'ppn_best.pth.tar' (model with best validation accuracy) will be generated in the folder 'exps/snapshots'
• --resume path to the model for recovering training
• -j number of workers for loading data
• --print-freq print frequency

*Training log ppn.log is uploaded to the folder exps/logs for the reference.

Testing

Run the following command to evaluate PPN on MPII validation set:

sh run_test.sh

or

CUDA_VISIBLE_DEVICES=0 python main.py --evaluate True --calc-map True --resume exps/snapshots/ppn_best.pth.tar

Run the following command to evaluate PPN on MPII testing set:

CUDA_VISIBLE_DEVICES=0 python main.py --evaluate True --resume exps/snapshots/ppn_best.pth.tar --eval-anno dataset/mpi/jsons/MPI_MP_TEST_annotations.json

In particular, results will be saved as a .mat file followed the official evaluation format of MPII Multi-Person Human Pose. An example is provided in exps/preds/mat_results/pred_keypoints_mpii_multi.mat.

Some configurable parameters in testing phase:

• --evaluate True for testing and false for training
• --resume path to the model for evaluation
• --calc-map calculate mAP or not
• --pred-path path to the mat file for saving the evaluation results
• --visualization visualize evaluation or not
• --vis-dir directory for saving the visualization results

The pretrained model and its performance (measured by mAP) on MPII validation set with this code:

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Avg.	Pretrained Model
PPN	94.0	90.9	81.2	74.1	77.1	73.4	67.5	79.7	GoogleDrive

*The Single-Person pose estimation model to refine Multi-Person pose estimation results will be released soon.

Citation

If you use our code/model in your work or find it is helpful, please cite the paper:

@inproceedings{nie2018ppn,
  title={Pose Partition Networks for Multi-Person Pose Estimation},
  author={Nie, Xuecheng and Feng, Jiashi and Xing, Junliang and Yan, Shuicheng},
  booktitle={ECCV},
  year={2018}
}