Skip to content

Latest commit

 

History

History
160 lines (109 loc) · 4.79 KB

README.md

File metadata and controls

160 lines (109 loc) · 4.79 KB

RetinaFace

Current Support Platform

RK3566, RK3568, RK3588, RK3562, RK1808, RV1109, RV1126

Model Source

The model used in this example comes from the following open source projects: https://github.com/biubug6/Pytorch_Retinaface

Download link:

RetinaFace_mobile320.onnx
RetinaFace_resnet50_320.onnx

Download with shell command:

cd model
./download_model.sh

(Optional) Convert original model to ONNX model

Refer to convert_to_onnx.py provided by the original project to convert the onnx file. The command is as follows:

python3 convert_to_onnx.py -m ./weights/mobilenet0.25_Final.pth --network mobile0.25 --long_side 320

Convert ONNX model to RKNN model

cd python
python convert.py <onnx_model> <TARGET_PLATFORM> <dtype(optional)> <output_rknn_path(optional)>
# such as: python convert.py ../model/RetinaFace_mobile320.onnx rk3566
# output model will be saved as ../model/RetinaFace.rknn

Description:

  • <onnx_model> should be the ONNX model path.
  • <TARGET_PLATFORM> could be specified as RK3562, RK3566, RK3568, RK3588, RK1808, RV1109, RV1126 according to board SOC version.
  • <dtype> is optional, could be specified as i8 or u8, i8/u8 means to do quantization, default is i8/u8.
  • <output_rknn_path> is optional, used to specify the saving path of the RKNN model, default save path is ../model/RetinaFace.rknn

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:

  • <TARGET_PLATFORM>: Specified as the NPU platform name. Such as 'rk3588'.
  • <rknn_model>: Specified as the model path.

Android Demo

Note: RK1808, RV1109, RV1126 does not support Android.

Compiling && Building

Modify the path of Android NDK in 'build-android.sh'.

For example,

export ANDROID_NDK_PATH=~/opt/toolchain/android-ndk-r18b

Then, run this script:

./build-android.sh -t <TARGET_PLATFORM> -a arm64-v8a -d RetinaFace

Please use the specific platform instead of <TARGET_PLATFORM> above.

Push all build output file to the board

Connect the USB port to PC, then push all demo files to the board.

adb root
adb push install/<TARGET_PLATFORM>_android_arm64-v8a/rknn_RetinaFace_demo/ /data/

Running

adb shell
cd /data/rknn_RetinaFace_demo/

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

Aarch64 Linux Demo

Compiling && Building

According to the target platform, modify the path of 'GCC_COMPILER' in 'build-linux.sh'.

export GCC_COMPILER=/opt/linux-x86/aarch64/gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu

Then, run the script:

./build-linux.sh  -t <TARGET_PLATFORM> -a aarch64 -d RetinaFace

Please use the specific platform instead of <TARGET_PLATFORM> above.

Push all build output file to the board

Push install/rknn_RetinaFace_demo_<TARGET_PLATFORM>_linux_aarch64 to the board,

  • If use adb via the EVB board:
adb push install/<TARGET_PLATFORM>_linux_aarch64/rknn_RetinaFace_demo /userdata/
  • For other boards, use the scp or other different approaches to push all files under install/rknn_RetinaFace_demo_<TARGET_PLATFORM>_linux_aarch64 to '/userdata'.

Please use the specific platform instead of above.

Running

adb shell
cd /userdata/rknn_RetinaFace_demo/

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

Note: Try searching the location of librga.so and add it to LD_LIBRARY_PATH if the librga.so is not found in the lib folder. Use the following command to add it to LD_LIBRARY_PATH.

export LD_LIBRARY_PATH=./lib:<LOCATION_LIBRGA>

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.

C DEMO Notes

  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.