README.md

TinyPose MNN Demo

This fold provides PicoDet+TinyPose inference code using Alibaba's MNN framework. Most of the implements in this fold are same as demo_ncnn.

Install MNN

Python library

Just run:

pip install MNN

C++ library

Please follow the official document to build MNN engine.

• Create picodet_m_416_coco.onnx and tinypose256.onnx example:

  modelName=picodet_m_416_coco
  # export model
  python tools/export_model.py \
          -c configs/picodet/${modelName}.yml \
          -o weights=${modelName}.pdparams \
          --output_dir=inference_model
  # convert to onnx
  paddle2onnx --model_dir inference_model/${modelName} \
          --model_filename model.pdmodel  \
          --params_filename model.pdiparams \
          --opset_version 11 \
          --save_file ${modelName}.onnx
  # onnxsim
  python -m onnxsim ${modelName}.onnx ${modelName}_processed.onnx

• Convert model example:

  python -m MNN.tools.mnnconvert -f ONNX --modelFile picodet-416.onnx --MNNModel picodet-416.mnn

Here are converted model picodet_m_416. tinypose256

Build

For C++ code, replace libMNN.so under ./mnn/lib with the one you just compiled, modify OpenCV path and MNN path at CMake file, and run

mkdir build && cd build
cmake ..
make

Note that a flag at main.cpp is used to control whether to show the detection result or save it into a fold.

#define __SAVE_RESULT__ // if defined save drawed results to ../results, else show it in windows

ARM Build

Prepare OpenCV library OpenCV_4_1.

mkdir third && cd third
wget https://paddle-inference-dist.bj.bcebos.com/opencv4.1.0.tar.gz
tar -zxvf opencv4.1.0.tar.gz
cd ..

mkdir build && cd build
cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-21 -DANDROID_TOOLCHAIN=gcc ..
make

Run

To detect images in a fold, run:

./tinypose-mnn [mode] [image_file]

param	detail
--mode	input mode，0:camera；1:image；2:video；3:benchmark
--image_file	input image path

for example:

./tinypose-mnn "1" "../imgs/test.jpg"

For speed benchmark:

./tinypose-mnn "3" "0"

Benchmark

Plateform: Kirin980 Model: tinypose256

param	Min(s)	Max(s)	Avg(s)
Thread=4	0.018	0.021	0.019
Thread=1	0.031	0.041	0.032

Reference

MNN