SGBM VisionWorks - Black Disparity Image

Hi, I’m trying to use the VisionWorks tool to calculate the disparity map of a stereo vision.
I read the documentation and I wrote a short code to get the map but the results are not good, so I decided to try with the images that OpenCV provides and do a test with those but the result is an all black image of disparity.
Sorry if I put external links but I could only upload one photo in the post, anyway they are the aloeL.jpg and aloeR.jpg images of the OpenCV stereo_match example


This is the result

These are the parameters of the function:

            0,                  // min_disparity
            64,                 // max_disparity
            8,                  // P1
            109,                // P2
            5,                  // sad
            0,                  // ct_win_size
            1,                  // hc_win_size
            31,                 // bt_clip_value
            32000,              // max_diff
            0,                  // uniqueness_ratio
            7,                 // scanlines_mask
            2);                //flags

For now I use OpenCV to show it, convert vx_image to Mat and then use these lines of code to show it:



Would you mind to share the complete code with us?

The disparity map should be a GPU buffer.
So you need to memcpy it back to CPU first.


Here, some pieces of code were in some functions recalled but I copied and connected the various parts in the right way should be like this.

cv::Mat frame_left,frame_right,frame_gray_left_,frame_gray_right_,disp,disp_8;

frame_left = cv::imread("aloeL.jpg",cv::IMREAD_COLOR);
frame_right = cv::imread("aloeR.jpg",cv::IMREAD_COLOR);


vx_image gray_left = nvx_cv::createVXImageFromCVMat(context_, frame_gray_left_);
vx_image gray_right = nvx_cv::createVXImageFromCVMat(context_, frame_gray_right_);

vx_uint32 width,height;

vx_image disparity = vxCreateImage(context_,width,height,VX_DF_IMAGE_S16);

nvx_cv::VXImageToCVMatMapper mapVXtoCV = nvx_cv::VXImageToCVMatMapper(disparity);    
disp = mapVXtoCV.getMat();



The default disparity is represented in Q11.4 fixed point format.
You will need a post-processing as below to convert disparity from fixed point to grayscale:

vx_image disparity_U8 = vxCreateImage(context, width,height, VX_DF_IMAGE_U8);
vx_int32 shift = 4;
vx_scalar s_shift = vxCreateScalar(context, VX_TYPE_INT32, &shift);
convert_depth_node_ = vxConvertDepthNode(main_graph_, disparity, disparity_U8, VX_CONVERT_POLICY_SATURATE, s_shift);


Thanks for your answer, I need time to rewrite the code using the graphs because I wasn’t using them in the code above, please wait for my news.

In the meantime I have these questions:

  • If I use other images that come from my stereo vision the disparity that comes back to me using the code above is not all black.(If you want I can upload the images).

  • Your solution uses a node to do the conversion, while I use a VisionWorks single call function in the code above, is there a way to do the conversion without using nodes?

I was going to set the code with graphs and nodes so these are just questions of curiosity.

Thanks for waiting, when I try to launch the graph with input these two images give me this error:

terminate called after throwing an instance of ‘std::runtime_error’
what(): vxProcessGraph(main_graph_) failure [status = -10] in file /home/nvidia/Desktop/stereoDepth/src/stereo_match_graph.cpp line 37
Aborted (core dumped)

debugging gives me an error on the input parameters of the SGBM node “VX_ERROR_INVALID_PARAMETERS”,
but I didn’t change anything in the VisionWorks code, not even the SGBM node parameters.
I tried the images captured by my stereo vision and I have no errors, I guess the problem is related to these two images, just out of curiosity can you understand why?

If there isn’t answer close topic .


Do you meet the error when using /usr/share/visionworks/sources/demos/stereo_matching/?
Or the error comes from your implementation?

If it’s customized, could you share the source with us for reproducing?


When I use the example in the VisionWorks folder I have no errors.
The code is above, where simply I read these two images with OpenCV I convert them in vx_image and I execute SGBM. Same code with other two images doesn’t give me errors, actually it doesn’t give me errors but the output is an all black screen.