ncnn-deep-head-pose

Simple implement inference deep head pose ncnn version with high performance and optimized resource. This project based on deep-head-pose project by Nataniel Ruiz. And detail paper in CVPR Workshop. I use Retinaface for face detection step.

Workflow

Re-build ncnn for arbitrary platform.

• Official ncnn document shown in detail how to use and build ncnn for arbitrary platform.
• And if you use my docker environment, i was build ncnn library inside docker environment with path: /home/ncnn_original/build it's contain ncnn shared static library and tools for convert and quantize ncnn models.

Convert models ncnn format.

• As original deep head pose project used Pytorch framwork. So, We need convert Pytorch model to ncnn model.
• In ncnn wiki had detailed this guide here. After convert pytorch to onnx format, you have to use ncnn build tools to convert onnx->ncnn. Inside my docker env, ready to use in /home/ncnn_original/build/tools/onnx
• Notice, Netron support to visualize network in intuitive easy to get input node and output node as specification of ncnn

Build run test

Run environment

• git clone https://github.com/docongminh/ncnn-deep-head-pose
• cd ncnn-deep-head-pose
• execute env: docker exec -it deep-head-pose bash
• cd to source mounted: cd /source
• cd to ncnn build library: cd /home/ncnn_original

Cmake project

• In project root inside docker: `mkdir -p build && cd build

• Cmake and build: cmake .. && make

• Run test: ./main

• Examples:

  

  

Note during develop project

• create extractor instant
• normalize image
• resize image
• This project in progress. So, it has many coding performance issues during develop process.
• Quantized model version will be update ASAP.

References

• deep-head-pose
• retinaface
• ncnn