YOLOv3 C++

Torch (C++) implementation of YOLOv3, which works on Windows,Mac, Linux.

Dependency

• cmake
• libTorch >= 1.3, or nightly
• OpenCV >= 4.0
• C++17
• Win10: vs2019, passed
• Linux:
  • g++ >=7.3.0
  • Pre-cxx11 ABI: passed
  • cxx11 ABI => error: libtorch.so: undefined reference to `lgammaf@GLIBC_2.23'
• MacOS: passed

Usage

For now, only support inference. Please use Darknet weights format as input.

If you need training your own model, try PyTorch-YOLOv3 and save your weights.

To test all jpg format images in a folder, use your yolo.custom.cfg and custom.weights, run the code below:

# yolov3 <config_file> <weights> <image_folder> 
yolov3.exe yolov3.cfg yolov3.weights images

Build

# before running this project, you have to edit two file
# -> edit the CMakeList.txt, set correct path to libtorch and OpenCV
# -> edit the yolov3/main.cpp at line 27. Set correct image format. 
cd path/to/YOLOv3
mkdir build
cd build
cmake ..
make # linux, mac
# if windows, open YOLOV3-app.sln and then build 

Performance

test code:

yolov3 models/yolov3.cfg models/yolov3.weights images

Results:

1. tested with CPU: Core i9

   • Windows
   • MAC: average time (682 ms/image).

2. tested with GPU: Tesla V100

   • Linux: average time (22 ms/image).

Features

Supports

• NMS
• Soft NMS
• Weighted NMS

TODO

• Support training
• ...

Credits

This repo is created based on the implementations below:
weixu000
PyTorch-YOLOv3
YOLO_v3_tutorial_from_scratch