YOLOv3_SpringEdition

YOLOv3 C++ Windows and Linux interface library. (Train,Detect both)

• Remove pthread,opencv dependency.
• You need only 1 files for YOLO deep-learning.
• Support windows, linux as same interface.

Do you want train YOLOv3 as double click? and detect using YOLOv3 as below?

YOLOv3 detector;
detector.Create("coco.weights", "coco.cfg", "coco.names");
cv::Mat img=cv::imread("a.jpg");
std::vector<BoxSE> boxes = detector.Detect(img, 0.5F);

• Then you've come to the right place.

1. Setup for train.

1.1. Train detector

You need only 2 files for train that are YOLOv3SE_Train.exe and cudnn64_7.dll on Windows. If you are on Linux, then you need only YOLOv3SE_Train. This files are in YOLOv3_SpringEdition/bin. or you can make it using build_windows.bat and build_linux.sh.

The requirement interface not changed. Same as pjreddie/darknet.

There is a example training directory Yolov3_SpringEdition_Train/. You can start training using above files.

Actually, all the interfaces are same with YOLOv2. So you can easily train your own data.

The YOLOv3SE_Train.exe's arguments are [option],[base directory],[data file path] and [cfg file path].

And YOLOv3SE_Train.exe is automatically choosing multi-gpu training. and select latest backup weights file.

Example : Yolov3_SpringEdition_Train/DetectorExample/

Sample directory structure with VOC2007 dataset.

┌ voc2007train
│  ├ 000012.jpg
│  ├ 000012.txt
│  ├ 000017.jpg
│  ...
│  └ 009961.txt
├ backup
├ yolov3_darknet53.cfg
├ voc2007.data
├ voc2007.names
├ train.txt
├ cudnn64_7.dll
└ YOLOv3SE_ Train.exe

Sample run argument with STL10 dataset.

"YOLOv3SE_Train.exe" detector . voc2007.data darknet53.cfg

1.2. Train classifier

Example : Yolov3_SpringEdition_Train/ClassifierExample/

Sample directory structure with STL10 dataset.

┌ stl10train
│  ├ airplane
│  ├ bird
│  ├ car
│  ...
│  └ truck
├ backup
├ resnext50.cfg
├ stl10.data
├ stl10.names
├ train.txt
├ cudnn64_7.dll
└ YOLOv3SE_ Train.exe

Sample run argument with STL10 dataset.

"YOLOv3SE_Train.exe" classifier . stl10.data resnext50.cfg

2. Setup for detect

Do not change "batch" and "subdivisions" in cfg file. It automatically read those values as 1 when the network is on testing mode.

Just include YOLOv3SE.h and use it. See YOLOv3_SpringEdition_Test/. You need only YOLOv3SE.h, libYOLOv3SE.dll and cudnn64_7.dll for detect.

1. Go to Yolov3_SpringEdition_Test

2. Run "download_cudnn64_7.dll.bat" if you're in Windows.

3. Download "voc2007valid", "yolov3_darknet53.weights" for detection.

Download "stl10valid", "resnext50_256_10000.weights" for classification.

4. Run VS solution or build.sh.

Reference

The class YOLOv3 that in YOLOv3SE.h has 3 methods.

void Create(std::string weights,std::string cfg,std::string names);

This method load trained model(weights), network configuration(cfg) and class naming file(names)*

• Parameter
  • weights : trained model path(e.g. "obj.weights")
  • cfg : network configuration file(e.g. "obj.cfg")
  • names : class naming file(e.g. "obj.names")

std::vector<BoxSE> Detect(cv::Mat img, float threshold);
std::vector<BoxSE> Detect(std::string file, float threshold);
std::vector<BoxSE> Detect(IplImage* img, float threshold);
int Classify(IplImage* img);
int Classify(cv::Mat img);
int Classify(std::string file);

This method is detecting objects or classify of file,cv::Mat or IplImage.

• Parameter
  • file : image file path
  • img : 3-channel image.
  • threshold : It removes predictive boxes if there score is less than threshold.

void Release();

Release loaded network.

Technical issue

Original YOLOv3(darknet) is linux version. And AlexeyAB already made YOLOv3 Windows version. But, his detection method is too slow on Windows. I don't know why exactly. Maybe it has bottleneck. So, I converted darknet(YOLOv3 only) again.

• YOLOv1 doesn't work.

change log

build_windows.bat and build_linux.sh will download automatically correct version of cudnn. and build as cmake.

Windows + 2080ti + CUDA 10.0 + cudnn7.3.1 + yolov3   = 44FPS

Windows + 1080ti + CUDA 8.0 + cudnn7.1 + yolov3      = 36FPS
Windows + 1080ti + CUDA 9.0 + cudnn7.1 + yolov3      = 36FPS

Windows + 1080   + CUDA 9.0 + cudnn7.1 + yolov3      = 27FPS
Windows + 1080   + CUDA 9.0 + cudnn7.1 + yolov3(spp) = 15FPS
Ubuntu  + 1080   + CUDA 8.0 + cudnn7.1 + yolov3      = 30FPS
Ubuntu  + 1080   + CUDA 9.0 + cudnn7.1 + yolov3      = 30FPS

Software requirement

• CMake
• CUDA 8.0 or 9.0(9.1 is not working) or 10.0
• OpenCV(for testing)
• Visual Studio 2013, 2015, 2017

Hardware requirement

• NVIDIA GPU