Skip to content

Latest commit

 

History

History
109 lines (91 loc) · 5.23 KB

README.md

File metadata and controls

109 lines (91 loc) · 5.23 KB
Raw

image

This version has been deprecated, see YoloV3-darknet!

This version is for you if you have to use C++ version faster-rcnn in the windows environment because the project needs.

Introduction

The windows C ++ version of faster-rcnn is based on two versions:  

一: Microsoft's windows-caffe

This version is Microsoft's windows caffe version    

二: D-X-Y's caffe-faster-rcnn version

This is the D-X-Y's Linux c ++ version of faster-rcnn

Features

  • After combining the above two, this is the C ++ version faster-rcnn base on windows, no python, support training and testing.
  • It takes about 60ms to test a image, under the GeForce GTX TITAN X , using the VGG16 model.

Compile（VS2013 x64 release mode）

  • Same as Microsoft Caffe, or refer to Blog
  • Here for convenience, I have put the cudnn files under the project, the cuda folder under the main directory.
  • If you want to configure cudnn, remember write the right cuda folder path in the "cuDnnPath" in the ".\Windows\CommonSettings.props" .

Train model

  • The steps to train the model are exactly the same as D-X-Y's caffe-faster-rcnn version.
  • You can read the script "caffe-master\examples\FRCNN\vgg16\train_frcnn.bat" for detail.
  • Of course, you can train your model under python version.

Test the model, calculate mAP and Recall

Run the script "caffe-master\examples\FRCNN\vgg16\test_frcnn.bat" for detail.

Use Faster-rcnn like OpenCV in VS2013

  • After compiling, all dependencies generated with faster-rcnn are under "caffe-master\Build\x64\Release".
  • Third-party dependencies, such as OpenCV, Glog, protobuf, etc. are in the "NugetPackages".
  • Just as VS2013 configures OpenCV, configure faster-rcnn as long as the include-files for faster-rcnn and third-party libraries are placed into the VC ++ directory of VS2013, the lib files are placed into the library directory, and linker-> input-> additional dependencies, input:

libboost_date_time-vc120-mt-1_59.lib
libboost_filesystem-vc120-mt-1_59.lib
libboost_system-vc120-mt-1_59.lib
libglog.lib
libcaffe.lib
gflags.lib
gflags_nothreads.lib
hdf5.lib
hdf5_hl.lib
libprotobuf.lib
libopenblas.dll.a
Shlwapi.lib
opencv_core2410.lib
opencv_highgui2410.lib
opencv_imgproc2410.lib
LevelDb.lib
lmdb.lib
opencv_video2410.lib
opencv_objdetect2410.lib
cublas.lib
cuda.lib
cublas_device.lib
cudart.lib
cudart_static.lib
curand.lib
kernel32.lib

  • Refer to the blog: http://blog.csdn.net/auto1993/article/details/70198435 for detail.
  • Because the third-party libraries are configured more trouble, for convenience, I packaged the NugetPackages files in the third-party library on Baidu network disk, the password : sj7f
  • There are three folders in the packaged folder: bin, include, lib, respectively, corresponding to the executable file, header files, library files.

Test images

After VS2013 configuration successful, you can write code in VS2013 to test the pictures. The detector was defined in file caffe-master\include\caffe\api\FRCNN\frcnn_api.hpp:

#include <caffe\api\FRCNN\frcnn_api.hpp>  //Detect head file  
#include "Register.h"           //This file is necessary used to register the relevant caffe layer 
using namespace std;  
using namespace cv;  
using namespace caffe::Frcnn;  
int main(){    
        Mat frame = imread("1.jpg");  //image  
	/* Initiaze the detector, the four parameters were:  
	1. network file   
	2. trained model file  
	3. config file  
	4. whether to open the GPU mode, default true  
	5. whether to ignore print log, default true
	*/
	FRCNN_API::Detector detect("VGG16.prototxt", "VGG16.caffemodel", "config_file.json", true, true);  
	vector<BBox<float> > boxes = detect.predict(frame);    // forward, detect results saved here 
	for (int i = 0; i < boxes.size(); i++)   //draw rects  
	     rectangle(frame, cv::Point(boxes[i][0], boxes[i][1]), cv::Point(boxes[i][2], boxes[i][3]), Scalar(0, 0, 255)); 
	imshow("demo", frame);  
        waitKey(0);  
	return 0;  
}

Register.h、VGG16.prototxt、VGG16.caffemodel、config_file.json files

  • Need to add Register.h to your code, otherwise it will complain that no registration related layer, I had put this file in the main project directory.
  • "VGG16.prototxt" -- a network description file under path "caffe-master\models\FRCNN\vgg16"
  • "VGG16.caffemodel" -- trained model file
  • "config_file.json" -- training and testing need to use the configuration file, it involves the number of target categories, NMS threshold, etc. under "caffe-master\examples\FRCNN\config"

others