Segmentation fault when using nvdspreprocess

[nam1012.2016]

Just to be clear,converted_buffer->dataPtr point to raw frame data or the cropped object data in nvdspreprocess?

[yuweiw]

@nam1012.2016

converted_buffer->dataPtr

It’s the Scaled & converted buffer to processing width/height.

[nam1012.2016]

char*  src_data = (char*) malloc(surf_params->dataSize);
      if(src_data == NULL) {
          g_print("Error: failed to malloc src_data \n");
          continue;
      }

cudaMemcpy((void*)src_data,
        (void*)surf_params->dataPtr,
        surf_params->dataSize,
        cudaMemcpyDeviceToHost);

cv::Mat frame(surf_params->height, surf_params->width,CV_8UC4 ,src_data, surf_params->pitch);
cv::Mat image_data(surf_params->height, surf_params->width,CV_8UC4 );
frame.copyTo(image_data);
cv::Mat outptut_image(surf_params->height, surf_params->width,CV_8UC3 );
cv::cvtColor(image_data,outptut_image, cv::COLOR_RGBA2BGR);
cv::imwrite("output.jpg",outptut_image);

Thanks a lot for your help, absolutely appreciate it. This code works and now i’m getting some proper image. But how can i use unified memory to avoid data copying?

[yuweiw]

@nam1012.2016

But how can i use unified memory to avoid data copying?

If you want to avoid data copying, you need to do some researches on how OpenCV uses cuda buffers yourself. We don’t have demo for this scenario yet.

[nam1012.2016]

Thanks for your support. I can now access frame with unified memory. Im closing this topic now