[QUESTION] NVCV_COLOR_RGB2GRAY ERROR

[lzjia-jia]

void saveCrop(const nvcv::Tensor& crop, const std::string& filename)
{
    auto dataCuda = crop.exportData<nvcv::TensorDataStridedCuda>();
    auto sizes    = crop.shape();          // [N, H, W, C]
    int N = sizes[0], H = sizes[1], W = sizes[2], C = sizes[3];

    std::cout << "Input image N: " << N << std::endl;
    std::cout << "Input image H: " << H << std::endl;
    std::cout << "Input image W: " << W << std::endl;
    std::cout << "Input image C: " << C << std::endl;

    std::vector<uint8_t> host(N * H * W * C);
    cudaMemcpy(host.data(), dataCuda->basePtr(),
            host.size() * sizeof(uint8_t), cudaMemcpyDeviceToHost);

    cv::Mat img(H, W, CV_8U, host.data()); 

    cv::imwrite(filename, img);
    
}

   int maxImageWidth  = extParams.maxImageWidth;
    int maxImageHeight = extParams.maxImageHeight;
    int maxChannels    = extParams.maxChannels;

    // tag: Create the cuda stream
    cudaStream_t stream;
    CHECK_CUDA_ERROR(cudaStreamCreate(&stream));


    nvcv::TensorDataStridedCuda::Buffer inBuf;
    inBuf.strides[3] = sizeof(uint8_t);
    inBuf.strides[2] = maxChannels * inBuf.strides[3];
    inBuf.strides[1] = maxImageWidth * inBuf.strides[2];
    inBuf.strides[0] = maxImageHeight * inBuf.strides[1];
    CHECK_CUDA_ERROR(cudaMallocAsync(&inBuf.basePtr, batchSize * inBuf.strides[0], stream));

    nvcv::Tensor::Requirements inReqs
        = nvcv::Tensor::CalcRequirements(batchSize, {maxImageWidth, maxImageHeight}, nvcv::FMT_RGB8);

    nvcv::TensorDataStridedCuda inData(nvcv::TensorShape{inReqs.shape, inReqs.rank, inReqs.layout},
                                        nvcv::DataType{inReqs.dtype}, inBuf);

    nvcv::Tensor inTensor = nvcv::TensorWrapData(inData);

    uint8_t             *gpuInput = reinterpret_cast<uint8_t *>(inBuf.basePtr);
    
    uint32_t             totalImages = batchSize;

    nvjpegOutputFormat_t outputFormat = NVJPEG_OUTPUT_RGBI;

    std::vector<int> w(batchSize);
    std::vector<int> h(batchSize);
    NvDecode(images_dir, batchSize, totalImages, outputFormat, gpuInput, w, h);
    int offset = 0;
    for(int i=0;i<batchSize;i++){
        int wOrigin = w[I];
        int hOrigin = h[I];

        nvcv::TensorDataStridedCuda::Buffer inputBuf;
            // 设置 NHWC 的 stride
        inputBuf.strides[3] = sizeof(uint8_t);          // C
        inputBuf.strides[2] = 3 * inputBuf.strides[3];  // W
        inputBuf.strides[1] = wOrigin * inputBuf.strides[2]; // H
        inputBuf.strides[0] = hOrigin * inputBuf.strides[1]; // N
        
        inputBuf.basePtr = reinterpret_cast<NVCVByte *>(gpuInput+offset);
        offset += wOrigin * hOrigin * 3;

        nvcv::Tensor::Requirements inReqs1
            = nvcv::Tensor::CalcRequirements(1, {wOrigin, hOrigin}, nvcv::FMT_RGB8);

        nvcv::TensorDataStridedCuda inData1(nvcv::TensorShape{inReqs1.shape, inReqs1.rank, inReqs1.layout},
                                            nvcv::DataType{inReqs1.dtype}, inputBuf);

        // Wrap 成 Tensor
        nvcv::Tensor inputTensor = nvcv::TensorWrapData(inData1);
         
       nvcv::Tensor gray(1, {w, h}, nvcv::FMT_U8);
       cvcuda::CvtColor cvtOp;
       cvtOp(stream, inputTensor, gray, NVCV_COLOR_RGB2GRAY);
      saveCrop(gray, "gray.jpeg");
}

cvtOp(stream, inputTensor, gray, NVCV_COLOR_RGB2GRAY);
Why does my grayscale image look like this after the conversion?

[justincdavis]

Hi @lzjia-jia , thank you for your interest in CV-CUDA!

Are you able to provide more information about your environment such as the CV-CUDA version, OS, and hardware?

My immediate thought would be that there is striding and/or row-alignment on the NVCV Tensor and when you copy the data you do not account for this and use a contiguous memcpy. If that is the case, it could cause an issue when viewing with cv::imwrite. Are you able to verify whether or not the tensor has striding or row-alignment prior to copying to host?