Failure to load Mediapipe's Face landmark model on DLPU

[Poufy]

We are trying to load the Mediapipe's Face landmark model on the DLPU using the Larod API but we fail to do so with the following error:

And running journalctl

We are however able to load different models like Mediapipe's Face detection model

To reproduce

We're following the same setup as the vdo-larod example, however, we're using CPP. Below is the function where we attempt to load the model

bool setupLarod(const char *modelFile, larodConnection **conn, larodModel **model)
{
    larodError *error = NULL;
    *conn = NULL;
    *model = NULL;
    int larodModelFd = -1;
    const larodDevice *dev = NULL;

    if (!larodConnect(conn, &error))
    {
        std::cout << "Could not connect to larod: " << error->msg << std::endl;
        goto error;
    }

    larodModelFd = open(modelFile, O_RDONLY);
    if (larodModelFd < 0)
    {
        std::cout << "Unable to open model file: " << modelFile << std::endl;
        goto error;
    }


    dev = larodGetDevice(*conn, "axis-a8-dlpu-tflite", 0, &error);
    if (!dev)
    {
        std::cout << "Unable to get device: " << error->msg << std::endl;
        goto error;
    }

    *model = larodLoadModel(*conn, larodModelFd, dev, LAROD_ACCESS_PRIVATE, "", NULL, &error);
    if (!*model)
    {
        std::cout << "Unable to load model: " << error->msg << std::endl;
        goto error;
    }

    close(larodModelFd);
    return true;

error:
    if (larodModelFd >= 0)
    {
        close(larodModelFd);
    }
    if (*conn)
    {
        larodDisconnect(conn, NULL);
    }
    larodClearError(&error);
    return false;
}

And we call it like so

        char modelFile[] = "/usr/local/packages/opencv_app/face_landmark.tflite";
        if (!setupLarod(modelFile, &larodConn, &model))
        {
            std::cout << "setupLarod failed" << std::endl;
            exit(1);
        }

Environment

• Axis device model: AXIS M5526-E PTZ Camera
• Axis device firmware version: 11.11.73

Additional context

When swapping the device to use the CPU instead, the model is loaded without any issues.

Is this a hardware restriction? Is that operations in the model aren't supported by the DLPU?

[pataxis]

Hi @Poufy , thanks for reaching out with your question.

We have a good documentation on computer vision with many details, please have a look there first if it can answer your question.