springkim / mmodcnn_springedition Goto Github PK

Max-margin object detection convolution neural networks.

• Remove dlib dependency.
• You need only 1 files for MMODCNN object detection.
• Support windows, linux as same interface.

MMODCNN detector;
detector.Create("vehicle.dat", "names.txt");
cv::Mat img=cv::imread("a.jpg");
std::vector<BoxSE> boxes = detector.Detect(img, 0.5F);

You need only 2 files for train that are MMODCNNSE_Train.exe and cudnn64_7.dll on Windows. If you are on Linux, then you need only MMODCNNSE_Train. This files are in build/Release after run cmake build.

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

The MMODCNN_Train.exe's arguments are [directory],[xml],[network name],[mini batch size] , [min detection width=40], [min detection height=40], [solver=sgd], [Cropper=350] and [min learning rate=0.0001].

MMODCNNSE_Train.exe . training.xml hello 32 40 40 adam 500

Just include MMODCNNSE.h and use it. See MMODCNN_SpringEdition_Test/. You need only MMODCNNSE.h, libMMODCNNSE.dll, opencv_world400.dll and cudnn64_7.dll for detect.

The class MMODCNN that in MMODCNNSE.h has 3 methods.

void Create(std::string dat, std::string names);

This method load trained model(weights) that has dat extension, and class naming file(names)

• Parameter
  • dat : trained model path(e.g. "vehicle.dat")
  • names : class naming file(e.g. "names.txt")

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

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.

This is dlib's MMODCNN wrapper. I added dynamic mini-batch loading module when training one step.

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

Windows + 1080ti + CUDA 10.0 + cudnn7.5      = 28FPS

• CMake
• CUDA 10.0
• OpenCV(for testing)(included in repository)
• Visual Studio 2013, 2015, 2017

• NVIDIA GPU