opencv cuda使用笔记

cuda拷贝

cuda+opencv yolov4

颜色转换：

cv::cuda::split

查看10*10像素

down_load用法

特别全的学习资料：

我想将OpenCV::cuda::GpuMat类型的数据给TensorRT GPU加速，请问有方法吗？ - 知乎

cuda拷贝

cv::Mat image = cv::imread("C:\\RegressionModelImages\\cat.png", cv::IMREAD_COLOR);
cv::cvtColor(image, image, cv::COLOR_BGR2RGB);

This can be converted to a GpuMat if necessary:

cv::cuda::GpuMat gpu_image(image.size(), image.type());

cv::Mat frame_cpu;cv::cuda::GpuMat d_frame_gpu;cv::Mat down_img;double t = (double)cv::getTickCount();frame_cpu=cv::imread("/home/tao/useful/faceboxes_tensorrt_cuda/cuda_opencv/test.jpg");cvtColor(frame_cpu,frame_cpu,CV_RGB2BGRA);frame_cpu.convertTo(frame_cpu, CV_32FC4,4);std::cout<<frame_cpu.channels()&lt;&lt;std::endl;d_frame_gpu.upload(frame_cpu);frame_pad(d_frame_gpu,d_frame_gpu);cv::cuda::resize(d_frame_gpu,d_frame_gpu,cv::Size(1024,1024));size_t inputWidth=1024;size_t inputHeight=1024;size_t outputWidth=1024;size_t outputHeight=1024;float* output;float* output_host;cudaMalloc((void**)&output,sizeof(float)*inputWidth*inputWidth*3);cudaPreImageNet((float4*)d_frame_gpu.pt<float4>(), inputWidth, inputHeight,output, outputWidth, outputHeight);cudaMemcpy(output_host,output,sizeof(float)*inputWidth*inputWidth*3,cudaMemcpyDeviceToHost);

使用OpenCV操作CUDA送检图像到TensorRT(SSD模型)_ikevin的专栏-CSDN博客

cuda+opencv yolov4

使用CUDA+OpenCV加速yolo v4性能_深度学习与计算机视觉-CSDN博客

颜色转换：

cv::cuda::resize(gpu_frame, gpu_frame, Size(960, 540), 0, 0, INTER_LINEAR); // convert to gray cv::cuda::GpuMat gpu_current; cv::cuda::cvtColor(gpu_frame, gpu_current, COLOR_BGR2GRAY);

cv::cuda::split

cv::cuda::split 使用_jacke121的专栏-CSDN博客

        cv::Mat image = cv::imread(imgpath + filelist[i]);//std::cout << "image:" << filelist[i] << std::endl;if (image.empty()) {std::cout << "读入图片为空，请检查路径！" << std::endl;system("pause");return -1;}/* cv::Mat image;cv::cvtColor(image_o, image, cv::COLOR_BGR2RGB);*/DWORD start2 = GetTickCount();cv::cuda::GpuMat imageGpu(image.cols, image.rows, CV_8UC3, cv::Scalar(0, 0, 0));//cv::cuda::GpuMat imageGpu;// (cv::Size(image.cols, image.rows), CV_8UC3);cv::cuda::GpuMat imageRGB(image.cols, image.rows, CV_32FC3, cv::Scalar(0, 0, 0));imageGpu.upload(image);imageGpu.convertTo(imageRGB, CV_32FC3, 1,0);

转cpu：

 cv::Mat dst;flt_image_out.download(dst);

    cv::Mat dst_gold;cv::cvtColor(src, dst_gold, cv::COLOR_BGR2RGB);

查看10*10像素

区别：一个3b，一个3f。

cv::Mat raw_image = cv::imread(image_name.c_str());for (int i = 0; i <10; i++){for (int j = 0; j <10; j++){std::cout << "(" <<(float)raw_image.at<cv::Vec3b>(i, j)[0]<< ", " <<(float)raw_image.at<cv::Vec3b>(i, j)[1]<< ", " <<(float)raw_image.at<cv::Vec3b>(i, j)[2]<< ") " << std::endl;std::cout << "(" <<(float)raw_image.at<cv::Vec3f>(i, j)[0] << ", " <<(float)raw_image.at<cv::Vec3f>(i, j)[1] << ", " <<(float)raw_image.at<cv::Vec3f>(i, j)[2] << ") " << std::endl;}}

查看全部像素：


void showMat(cv::Mat &img){std::cout << "图像元素查看 ： " << img.rows<<"," << img.cols <<std::endl;for (int i = 0; i<img.rows; i++){for (int j = 0; j < img.cols; j++){std::cout<< "("<<(float)img.at<cv::Vec3f>(i, j)[0]<< ", " <<(float)img.at<cv::Vec3f>(i, j)[1]<< ", " <<(float)img.at<cv::Vec3f>(i, j)[2]<< ") " << std::endl;//(int)img.at<cv::Vec3b>(i, j)[0]<< ", " <<//(int)img.at<cv::Vec3b>(i, j)[1]<< ", " <<//(int)img.at<cv::Vec3b>(i, j)[2]<< ") " << std::endl;}}}
}

cuda方式操作：

OpenCV+CUDA 遍历cv::Mat笔记_YaoJiawei329的博客-CSDN博客

down_load用法

比较慢

        DWORD start2 = GetTickCount();cv::cuda::GpuMat imageGpu(cv::Size(image.cols, image.rows), CV_8UC3);cv::cuda::GpuMat imageRGB;imageGpu.upload(image);imageGpu.convertTo(imageRGB, CV_32FC3, 1,0);//std::cout << "dst:" << errMsg << std::endl;cv::Mat dst;imageRGB.download(dst);DWORD end2 = GetTickCount();std::cout << "gpu img time：" << end2 - start2 << "ms " << image.cols << " " << image.rows << std::endl;

gpu间拷贝：

https://github.com/LiPorkLi/BatteryFilmCheck/blob/c048163ac0d10cb7db2e7506b7e0c8d8f8a96a20/ProtoParams/Alogrithm/BoudarySearch.cpp

cudaMemcpy(SrcDstImg->data + iStep + iS, SrcDstImg->data + iStep + m_vecvecBoundary[i][j].first, sizeof(uchar)*iLength, cudaMemcpyDeviceToDevice);//memcpy(SrcDstImg->data + iStep + iS, SrcDstImg->data + iStep + m_vecvecBoundary[i][j].first, sizeof(uchar)*iLength);iE = m_vecvecBoundary[i][j].second + iLength;if (iE > SrcDstImg->cols){iE = SrcDstImg->cols - 1;iLength = iE - m_vecvecBoundary[i][j].second + 1;}cudaMemcpy(SrcDstImg->data + iStep + m_vecvecBoundary[i][j].second, SrcDstImg->data + iStep + (m_vecvecBoundary[i][j].second - iLength), sizeof(uchar)*iLength, cudaMemcpyDeviceToDevice);

reshape：

reshape完毕后还是二维向量，无法变成一维向量。

 cv::cuda:GPUMat input;cv::cuda:GPUMat input3;
input3 = input.reshape(1,1);