(valid for personal testing) compilation guide of paddedetection on Jetson

This document is in NVIDIA Jetson Xavier NX (Jetpack The version is 4.4.1) Upper Department PaddleDetection Of c++ Reasoning course .
Jetpack 4.4.1 Your environment is configured as ：
CUDA 10.2
cuDNN 8.8.0
1. Get deployment code
I downloaded is 2.3 edition .
The download link is as follows ：https://github.com/PaddlePaddle/PaddleDetection.git
2. download paddle_inference Forecast base , The download link is as follows ：https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html

The contents of the extracted folder are as follows ;
paddle_inference
├── paddle # paddle Core libraries and header files
├── third_party # Third parties rely on Libraries and header files
└── version.txt # Version and compilation information
3. Get into PaddleDetection-release-2.3\deploy\cpp\scripts In file
compile cmake Command on scripts/build.sh in , adopt vim Modify the main parameters , The modified example is as follows ：

#  Whether to use GPU( I.e. whether to use  CUDA)
WITH_GPU=ON

#  Whether to use MKL or openblas,TX2 I need to set to OFF
WITH_MKL=ON

#  Is it integrated  TensorRT( only WITH_GPU=ON  It works )
WITH_TENSORRT=ON

# paddle  Forecast base lib name , Due to different platforms and different versions of prediction Libraries lib Different names , Please check the downloaded forecast Library `paddle_inference/lib/` Under the folder `lib` The name of 
PADDLE_LIB_NAME=libpaddle_inference

# TensorRT  Of include route 
TENSORRT_INC_DIR=/usr/include/aarch64-linux-gnu

# TensorRT  Of lib route 
TENSORRT_LIB_DIR=/usr/lib/aarch64-linux-gnu

# Paddle  Prediction library path （ Downloaded in the previous step paddle_inference The path after decompression ）
PADDLE_DIR=/path/to/paddle_inference

# CUDA  Of  lib  route 
CUDA_LIB=/usr/local/cuda-10.2/lib64

# CUDNN  Of  lib  route 
CUDNN_LIB=/usr/lib/aarch64-linux-gnu/

#  Whether the key point model prediction function is enabled 
WITH_KEYPOINT=ON

#  Whether to enable the tracking model prediction function 
WITH_MOT=OFF

MACHINE_TYPE=`uname -m`
echo "MACHINE_TYPE: "${
    MACHINE_TYPE}


if [ "$MACHINE_TYPE" = "x86_64" ]
then
  echo "set OPENCV_DIR for x86_64"
  # linux The system downloads the precompiled opencv
  mkdir -p $(pwd)/deps && cd $(pwd)/deps
  wget -c https://paddledet.bj.bcebos.com/data/opencv-3.4.16_gcc8.2_ffmpeg.tar.gz
  tar -xvf opencv-3.4.16_gcc8.2_ffmpeg.tar.gz && cd ..

  # set OPENCV_DIR
  OPENCV_DIR=$(pwd)/deps/opencv-3.4.16_gcc8.2_ffmpeg

elif [ "$MACHINE_TYPE" = "aarch64" ]
then
  echo "set OPENCV_DIR for aarch64"
  # TX2 The platform downloads the precompiled through the following command opencv
  mkdir -p $(pwd)/deps && cd $(pwd)/deps
  wget -c https://bj.bcebos.com/v1/paddledet/data/TX2_JetPack4.3_opencv_3.4.6_gcc7.5.0.tar.gz
  tar -xvf TX2_JetPack4.3_opencv_3.4.6_gcc7.5.0.tar.gz && cd ..

  # set OPENCV_DIR
  OPENCV_DIR=$(pwd)/deps/TX2_JetPack4.3_opencv_3.4.6_gcc7.5.0/

else
  echo "Please set OPENCV_DIR manually"
fi

echo "OPENCV_DIR: "$OPENCV_DIR

#  There is no need to change the following 
rm -rf build
mkdir -p build
cd build
cmake .. \
    -DWITH_GPU=${
    WITH_GPU} \
    -DWITH_MKL=${
    WITH_MKL} \
    -DWITH_TENSORRT=${
    WITH_TENSORRT} \
    -DTENSORRT_LIB_DIR=${
    TENSORRT_LIB_DIR} \
    -DTENSORRT_INC_DIR=${
    TENSORRT_INC_DIR} \
    -DPADDLE_DIR=${
    PADDLE_DIR} \
    -DWITH_STATIC_LIB=${
    WITH_STATIC_LIB} \
    -DCUDA_LIB=${
    CUDA_LIB} \
    -DCUDNN_LIB=${
    CUDNN_LIB} \
    -DOPENCV_DIR=${
    OPENCV_DIR} \
    -DPADDLE_LIB_NAME=${
    PADDLE_LIB_NAME} \
    -DWITH_KEYPOINT=${
    WITH_KEYPOINT} \
    -DWITH_MOT=${
    WITH_MOT}

make
echo "make finished!"

After modifying the script and setting the main parameters , perform build Script ：

sh ./scripts/build.sh

After execution build After script , Will be in deploy/cpp Generate... In file build Folder , The compiled file is in build In file .
4. Prediction and visualization
After successful compilation , The prediction procedure is build/main, The main command parameters are described as follows ：

 Parameter description ：
--model_dir	 The path of the exported detection and prediction model 
--model_dir_keypoint	Option
--image_file	 Image file path to predict 
--image_dir	 Picture folder path to predict 
--video_file	 Video file path to predict 
--camera_id	Option
--device	 Device at run time , Can choose CPU/GPU/XPU, The default is CPU
--gpu_id	 Specify the... For reasoning GPU device id( The default value is 0)
--run_mode	 Use GPU when , The default is paddle,  Optional （paddle/trt_fp32/trt_fp16/trt_int8）
--batch_size	 Test the prediction of the model batch size, In the specified image_dir Effective when 
--batch_size_keypoint	 Prediction of key point model batch size, The default is 8
--run_benchmark	 Whether to repeat the prediction benchmark velocity measurement  ｜
--output_dir	 The folder where the output picture is located ,  The default is output ｜
--use_mkldnn	CPU Whether it is enabled in prediction MKLDNN Speed up 
--cpu_threads	 Set up cpu Number of threads , The default is 1
--use_dark	 Whether the key point model output prediction uses DarkPose post-processing , The default is true

Examples are as follows ：

./main --model_dir=/path/to/yolov3_darknet --image_file=/path/to/test.jpg --device=GPU