Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time
February 10, 2019 18:27
March 5, 2019 08:24
January 17, 2019 08:38
December 15, 2019 11:17

Fast Human Pose Estimation Pytorch

This is an unoffical implemention for paper Fast Human Pose Estimation, Feng Zhang, Xiatian Zhu, Mao Ye. Most of code comes from pytorch implementation for stacked hourglass network pytorch-pose. In this repo, we followed Fast Pose Distillation approach proposed by Fast Human Pose Estimation to improve accuracy of a lightweight network. We first trained a deep teacher network (stacks=8, standard convolution, 88.33@Mpii pckh), and used it to teach a student network (stacks=2, depthwise convolution, 84.69%@Mpii pckh). Our experiment shows 0.7% gain from knowledge distillation.

I benchmarked the light student model hg_s2_b1_mobile_fpd and got 43fps on i7-8700K via OpenVino. Details can be found from Fast_Stacked_Hourglass_Network_OpenVino

Please check the offical implementation by fast-human-pose-estimation.pytorch

Update at Feb 2019

  • Model trained by using extra unlabeled images uploaded, hg_s2_b1_mobile_fpd_unlabeled shows 0.28% extra gain from knowledge transfered from teacher on unlabeled data.
  • The key idea is inserting unlabeled images into mpii dataset. For unlabeled samples, loss comes from difference b/w teacher and student. For labeled samples, loss is the sum of teacher-vs-student and student-vs-groundtruth.


hg_s8_b1: teacher model, hg_s2_b1_mobile:student model, hg_s2_b1_mobile_kd: student model trained with FPD. hg_s2_b1_mobile_fpd_unlabeled: student model trained with FPD with extral unlabeled samples.

Model in_res featrues # of Weights Head Shoulder Elbow Wrist Hip Knee Ankle Mean GFlops Link
hg_s8_b1 256 128 25.59m 96.59 95.35 89.38 84.15 88.70 83.98 79.59 88.33 28 GoogleDrive
hg_s2_b1_mobile 256 128 2.31m 95.80 93.61 85.50 79.63 86.13 77.82 73.62 84.69 3.2 GoogleDrive
hg_s2_b1_mobile_fpd 256 128 2.31m 95.67 94.07 86.31 79.68 86.00 79.67 75.51 85.41 3.2 GoogleDrive
hg_s2_b1_mobile_fpd_unlabeled 256 128 2.31m 95.94 94.11 87.18 80.69 87.03 79.17 74.82 85.69 3.2 GoogleDrive


  1. Create a virtualenv

    virtualenv -p /usr/bin/python2.7 pose_venv
  2. Clone the repository with submodule

    git clone --recursive
  3. Install all dependencies in virtualenv

    source posevenv/bin/activate
    pip install -r requirements.txt
  4. Create a symbolic link to the images directory of the MPII dataset:

    ln -s PATH_TO_MPII_IMAGES_DIR data/mpii/images
  5. Disable cudnn for batchnorm layer to solve bug in pytorch0.4.0

    sed -i "1194s/torch\.backends\.cudnn\.enabled/False/g" ./pose_venv/lib/python2.7/site-packages/torch/nn/

Quick Demo

  • Download pre-trained modelhg_s2_b1_mobile_fpd) and save it to somewhere, i.e checkpoint/mpii_hg_s2_b1_mobile_fpd/
  • Run demo on sample image
python tools/ -a hg -s 2 -b 1 --mobile True --checkpoint checkpoint/mpii_hg_s2_b1_mobile_fpd/model_best.pth.tar --in_res 256 --device cuda 
  • You will see the detected keypoints drawn on image on your screen

Training teacher network

  • In our experiments, we used stack=8 input resolution=256 as teacher network
python example/ -a hg --stacks 8 --blocks 1 --checkpoint checkpoint/hg_s8_b1/ 
  • Run evaluation to get val score.
python tools/ -a hg --stacks 8 --blocks 1 --checkpoint checkpoint/hg_s8_b1/preds_best.mat 

Training with Knowledge Distillation

  • Download teacher model's checkpoint or you can train from scratch. In our experiments, we used hg_s8_b1 as teacher.

  • Train student network with knowledge distillation from teacher

python example/ -a hg --stacks 2 --blocks 1 --checkpoint checkpoint/hg_s2_b1_mobile/ mobile=True --teacher_stack 8 --teacher_checkpoint 


Run evaluation to generate mat file

python example/ -a hg --stacks 2 --blocks 1 --checkpoint checkpoint/hg_s2_b1/ --resume checkpoint/hg_s2_b1/model_best.pth.tar -e
  • --resume_checkpoint is the checkpoint want to evaluate

Run tools/ to get val score

Export pytorch checkpoint to onnx

python tools/ -a hg -s 2 -b 1 --num-classes 16 --mobile True --in_res 256  --checkpoint checkpoint/model_best.pth.tar 
--out_onnx checkpoint/model_best.onnx 


  • --checkpoint is the checkpoint want to export
  • --out_onnx is the exported onnx file