ElectronBotCli/retinaface_rknn

RetinaFace

Table of contents

• 1. Description
• 2. Current Support Platform
• 3. Pretrained Model
• 4. Convert to RKNN
• 5. Python Demo
• 6. Android Demo
  • 6.1 Compile and Build
  • 6.2 Push demo files to device
  • 6.3 Run demo
• 7. Linux Demo
  • 7.1 Compile && Build
  • 7.2 Push demo files to device
  • 7.3 Run demo
• 8. Expected Results

1. Description

The model used in this example comes from the following open source projects:

https://github.com/biubug6/Pytorch_Retinaface

2. Current Support Platform

RK3562, RK3566, RK3568, RK3576, RK3588, RV1126B, RV1109, RV1126, RK1808, RK3399PRO

3. Pretrained Model

Download link:

RetinaFace_mobile320.onnx
RetinaFace_resnet50_320.onnx

Download with shell command:

cd model
./download_model.sh

4. Convert to RKNN

Usage:

cd python
python convert.py <onnx_model> <TARGET_PLATFORM> <dtype(optional)> <output_rknn_path(optional)>

# such as: 
python convert.py ../model/RetinaFace_mobile320.onnx rk3588
# output model will be saved as ../model/RetinaFace_mobile320.rknn

Description:

• <onnx_model>: Specify ONNX model path.
• <TARGET_PLATFORM>: Specify NPU platform name. Support Platform refer here.
• <dtype>(optional): Specify as i8 or fp. i8 for doing quantization, fp for no quantization. Default is i8.
• <output_rknn_path>(optional): Specify save path for the RKNN model, default save in the same directory as ONNX model with name RetinaFace_mobile320.rknn

5. Python Demo

Usage:

cd python
# Inference with RKNN model
python RetinaFace.py --model_path <rknn_model> --target <TARGET_PLATFORM>
# The inference result will be saved as the image result.jpg.

Description:

• : Specified as the NPU platform name. Such as 'rk3588'.
• : Specified as the model path.

6. Android Demo

6.1 Compile and Build

Usage:

# go back to the rknn_model_zoo root directory
cd ../../
export ANDROID_NDK_PATH=<android_ndk_path>

./build-android.sh -t <TARGET_PLATFORM> -a <ARCH> -d RetinaFace

# such as 
./build-android.sh -t rk3588 -a arm64-v8a -d RetinaFace

Description:

• <android_ndk_path>: Specify Android NDK path.
• <TARGET_PLATFORM>: Specify NPU platform name. Support Platform refer here.
• <ARCH>: Specify device system architecture. To query device architecture, refer to the following command:

  # Query architecture. For Android, ['arm64-v8a' or 'armeabi-v7a'] should shown in log.
  adb shell cat /proc/version
  

6.2 Push demo files to device

With device connected via USB port, push demo files to devices:

adb root
adb remount
adb push install/<TARGET_PLATFORM>_android_<ARCH>/rknn_RetinaFace_demo/ /data/

6.3 Run demo

adb shell
cd /data/rknn_RetinaFace_demo

export LD_LIBRARY_PATH=./lib
./rknn_retinaface_demo model/RetinaFace_mobile320.rknn model/test.jpg

• After running, the result was saved as result.jpg. To check the result on host PC, pull back result referring to the following command:

  adb pull /data/rknn_RetinaFace_demo/result.jpg
  

7. Linux Demo

7.1 Compile and Build

usage

# go back to the rknn_model_zoo root directory
cd ../../

# if GCC_COMPILER not found while building, please set GCC_COMPILER path
(optional)export GCC_COMPILER=<GCC_COMPILER_PATH>

./build-linux.sh -t <TARGET_PLATFORM> -a <ARCH> -d RetinaFace

# such as 
./build-linux.sh -t rk3588 -a aarch64 -d RetinaFace

Description:

• <GCC_COMPILER_PATH>: Specified as GCC_COMPILER path.

  export GCC_COMPILER=~/opt/arm-rockchip830-linux-uclibcgnueabihf/bin/arm-rockchip830-linux-uclibcgnueabihf
  

• <TARGET_PLATFORM> : Specify NPU platform name. Support Platform refer here.

• <ARCH>: Specify device system architecture. To query device architecture, refer to the following command:

  # Query architecture. For Linux, ['aarch64' or 'armhf'] should shown in log.
  adb shell cat /proc/version
  

7.2 Push demo files to device

• If device connected via USB port, push demo files to devices:

adb push install/<TARGET_PLATFORM>_linux_<ARCH>/rknn_RetinaFace_demo/ /userdata/

• For other boards, use scp or other approaches to push all files under install/<TARGET_PLATFORM>_linux_<ARCH>/rknn_RetinaFace_demo/ to userdata.

7.3 Run demo

adb shell
cd /userdata/rknn_RetinaFace_demo

export LD_LIBRARY_PATH=./lib
./rknn_retinaface_demo model/RetinaFace_mobile320.rknn model/test.jpg

• After running, the result was saved as result.jpg. To check the result on host PC, pull back result referring to the following command:

  adb pull /userdata/rknn_RetinaFace_demo/result.jpg
  

• Note:

  1. For the generation of BOX_PRIORS_320 and BOX_PRIORS_640 in C demo post-processing, please refer to the PriorBox function in python/RetinaFace.py. This function aims to pre-generate the anchors box parameters. In order to speed up the demo post-processing, the C code directly generates the array.
  2. In C demo post-processing, num_priors is the number of anchor boxes. When the model shape is 320x320, num_priors is 4200. When the model shape is 640x640, num_priors is 16800.

8. Expected Results

This example will print the labels and corresponding scores of the test image detect results, as follows:

face @(302 76 476 300) score=0.999512

• Note: Different platforms, different versions of tools and drivers may have slightly different results.