Fast Human Pose Estimation Pytorch

This ia 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

Results

hg_s8_b1: teacher model, hg_s2_b1_mobile:student model, hg_s2_b1_mobile_kd: student model trained with FPD.

Model	in_res	featrues	# of Weights	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Mean	Link
hg_s8_b1	256	128	25.59m	96.59	95.35	89.38	84.15	88.70	83.98	79.59	88.33	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	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	GoogleDrive

Installation

1. Create a virtualenv

   virtualenv -p /usr/bin/python2.7 pose_venv
   

2. Clone the repository with submodule

git clone --recursive https://github.com/yuanyuanli85/Fast_Human_Pose_Estimation_Pytorch.git

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/functional.py
   

Training teacher network

• In our experiments, we used stack=8 input resolution=256 as teacher network

python example/mpii.py -a hg --stacks 8 --blocks 1 --checkpoint checkpoint/hg_s8_b1/ 

• Run evaluation to get val score.

python tools/mpii.py -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/mpii_kd.py -a hg --stacks 2 --blocks 1 --checkpoint checkpoint/hg_s2_b1_mobile/ mobile=True --teacher_stack 8 --teacher_checkpoint 
checkpoint/hg_s8_b1/model_best.pth.tar  

Evaluation

Run evaluation to generate mat file

python example/mpii.py -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/eval_PCKh.py to get val score

Export pytorch checkpoint to onnx

python tools/mpii_export_to_onxx.py -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 

Here

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

Reference

• OpenVino: https://github.com/opencv/dldt
• Pytorch-pose: https://github.com/yuanyuanli85/pytorch-pose
• Fast Human Pose Estimation https://github.com/yuanyuanli85/Fast_Human_Pose_Estimation_Pytorch