/libtorch-yolov3

libtorch1.6-yolov3 implementation, modified from https://github.com/walktree/libtorch-yolov3

Primary LanguageC++

libtorch-yolov3

A Libtorch implementation of the YOLO v3 object detection algorithm, written with pure C++, modified from https://github.com/walktree/libtorch-yolov3.

It can be seen the changes in issue: walktree/libtorch-yolov3#52 (comment)

I have successfully tested in Win10 using Visual Studio 2017 and Visual Studio 2019.

Requirements

  1. >=LibTorch v1.6.0 (Tested on 1.6,1.8 and 1.13)

    Download from https://download.pytorch.org/libtorch/nightly/cpu/libtorch-win-shared-with-deps-latest.zip （==Make sure that you have the download the windows version)==

    Add "${LIBTORCH}/lib" to system path.

  2. CUDA(Optional)

  3. OpenCV4.6 (Sugest using windows pre-build package)

    Add opencv/build/x64/vc15/bin to Windows system PATH (Support Visual Studio 2019)

  4. Git Bash or Cmder

To compile

  1. Cmake 3.15
  2. Visual Studio 2017 (VC 15) or Visual Studio 2019 (VC 16)
$ mkdir build && cd build
$ cmake -G "Visual Studio 15 2017 Win64" -T host=x64 -DCMAKE_PREFIX_PATH="your libtorch root" -DOpenCV_DIR="your opencv root" ..

For VS 2019

# $ cmake -G "Visual Studio 16 2019" -T host=x64 .. -DCMAKE_PREFIX_PATH="your libtorch root" -DOpenCV_DIR="your opencv root" ..

For example:

$ cmake -G "Visual Studio 16 2019" -T host=x64 .. -DCMAKE_PREFIX_PATH="E:\DeepLearning\libtorch" -DOpenCV_DIR="E:\ScienceComputing\opencv\build" ..
  1. Finally compile, be aware that we should use Release because we have used the released version of opencv
$ cmake --build . --config Release -j 3

your libtorch root is like E:\python\pytorch\libtorch and your opencv root is like "E:\ScientificComputing\opencv\build"(which has OpenCVConfig.cmake file). Finally libtorch1.6-yolov3\build\Release\yolo-app.exe is generated.

Running the detector

The first thing you need to do is to get the weights file for yolov3:

cd models
wget https://pjreddie.com/media/files/yolov3.weights

By default, the program will load yolov3 cfg and weights in model directory. It can be changed manually in main.cpp in line 32 and line 39 (Don't forget to recompile!).

Copy all .dll file from libtorch/lib to libtorch1.6-yolov3/build, then open git bash or cmder and execute:

$ cd libtorch1.6-yolov3/build
$ ./Release/yolo-app.exe ../imgs/person.jpg

The output result should be like this:

loading weight ...
weight loaded ...
start to inference ...
inference taken : 980 ms
3 objects found
Done

If you encounted some errors like:

  • error while loading shared libraries: torch_cpu.dll: cannot open shared object file: No such file or directory, be sure that you have put the libtorch/lib in your system path, or you can copy the dependent file to the same directory as yolo-app.exe is in.
  • error while loading shared libraries: opencv_world460.dll: cannot open shared object file: No such file or directory, be sure that you have put the opencv/build/x64/vc15/bin in your system path, or you can copyopencv_world460.dll to the same directory as yolo-app.exe is in.

The output image can be found in libtorch1.6-yolov3/build/out-det.jpg detection-output1 detection-output2

Change logs

  • 2022.11.15 Rename libtorch1.6-yolov3 to libtorch-yolov3