NCNN_SinglePoseEstimation

Introduction

The project mainly provides pre-trained model (trained from MPII dataset by Pytorch) to estimate the individual human posture through NCNN.

Installation

the related package:

1. g++ Cmake Opencv (C++)
2. Protobuf (Optional )
3. download model from BaiduCloud (Password：yp19) (FP32) , BaiduCloud (Password：22qe) (INT8)
4. download INT8 quantization Dataset ( MPII val Dataset) from BaiduCloud (Password：fbma) (Optional )

NOTE: Protobuf version must match Cmake version (In this project, Protobuf is 3.11.4 Cmake is 3.10.2)

git clone https://github.com/ZtoYtoQ/NCNN_SinglePoseEstimation.git

cd {your clone root-dir}
mkdir build
cd build
cmake ..
make -j2

Run

cd ..   ## return  your clone root-dir
cp demo.jpg build/examples/
cp {your model download path /hg_pose.bin}  build/examples/
cp {your model download path /hg_pose.param}  build/examples/
cd build/examples 
./hgpose demo.jpg

INT8 Quantization

cd {your root-dir}
cp {your model download path /hg_pose.bin}  build/tools/quantize/
cp {your model download path /hg_pose.param}  build/tools/quantize/
mv {your INT8 quantization Dataset path}  build/tools/quantize/
cd build/tools/quantize/
### Create model quantization table ###
./ncnn2table --param=hg_pose.param  --bin=hg_pose.bin --images=MPII_val_image/  --output=hg_pose.table  --thread=2 --mean=112,113,110 --norm=0.004,0.004,0.004

### Quantization mdoel ###
./ncnn2int8 hg_pose.param hg_pose.bin hg_pose-int8.param hg_pose-int8.bin hg_pose.table

### Get int8 model  hg_pose-int8.param hg_pose-int8.bin ###

Task Lists

• Adding INT8 Quantization、INT8 model
• building on Android
• Providing training code
• Training smaller model
• building on other platforms(Jetson、Raspberry...)
• More....

Reference

NCNN：https://github.com/Tencent/ncnn

License

MIT