AcceleratedVision

This lib is for accelerating general computer vision algorithms, and fined-tuned.

Supported Platforms

• Windows
• Linux(Coming)

Now, it fits for Windows only, but it's easy to port to supporting Linux. My new toy jetson tx2 just didn't work, back to distributor now, I have to wait for a while.

Dependencies

1. Cmake 3.10+
2. OpenCV 3.x(it doesn't need to compile, just download it and extract to folder C:\Program Files\opencv
3. CUDA SDK(8.0 or later)
4. VS 2015 or later Tips: Your video card should be capable of compute capability 2.1 at least.

Build

# get the repo, and make sure those dependencies is installed.
git clone https://github.com/Dengjianping/AcceleratedVision.git

# configure
mkdir build && cd build
cmake -G "Visual Studio 14 Win64 ..

# build, a dll and lib files will be generated at build\Release
msbuild /t:rebuild /p:Configuration=Release /p:Platform=x64 accel_vision.sln

Usage

Copy that folder named include to your project folder, as well as the dll and lib files you just compiled.

#include "include\cuda_color_converter.h"

using namespace std;
using namespace cv;

#prgram comment(lib,"accel_vision.lib")

int main(void)
{
    string path = "2028.jpg";
    Mat img = imread(path), result;

    cudaGray(img, result); // gray this image

    string title = "CUDA";
    namedWindow(title);

    imshow(title, result);
    waitKey(0);

    return 0;
}

APIs List

API Name	Description
cudaThreshold	threshold a image
cudaGray	RGB to gray
cudaHSV	RGB color to HSV
cudaSplit	split a 3-ch image into 1-ch image
cudaMerge	merge three 1-ch images as 3-ch image
cudaErode	erode image
cudaDilation	dilate image
cudaLUT	look up table
cudaGaussianBlut	blur image by gaussian kernel
cudaLalace	laplace operator to extract edges
cudaSobel	sobel operator to extract edges
cudaHistogram	histogram for a image
cudaEqualizeHist	equalize a image
cudaADD/OR/Sqrt	some mathematical operations
cudaAdjust	adjust image intensity values or colormap
cudaAntisotropy	reduce image noise without removing significant parts of the image content, see reference
cudaGammaCorrection	increases or decreases gamma value of input image
cudaSum	the sum of all pixels/single image
cudaMean	the mean value of all pixels/single image
cudaIntegral	get a integral image for a single image
cudaTranspose	transpose a image/single image
cudaGaussianBackProjection	gaussian back projection, try test/images/2048_RGB.jpg as input, model.jpg as model
updating	---

Performance

I just test these api on my laptop(Thinkpad T420), which equips with CPU I5-2520M video card Nvidia NVS 4200M, that has only 1 SM, 48 CUDA cores, compute capability 2.1.

The size of testing image is 2048 * 2048. BTW, time cost only means the kernel executing time on gpu device, excluding data copying time on operations host-to-device and device-to-host.

API Name	Time Cost
cudaGray	1.85ms(RGB to gray)
cudaHSV	5.67ms(RGB to HSV)
cudaGaussianBlur	3.81ms(5*5 kernel size, single channel)

More details see the excel.