grabcuts

Overview #

Image segmentation using GrabCut algorithm.

Expected Outputs #

Prerequisite #

• OpenCV Viewer

Tutorial #

const int MAX_FOV_DEPTH_MM = 700;
const int DISPLAY_MIN_Z = 300;
const int DISPLAY_MAX_Z = 3000;

const int MAX_FOV_DEPTH_MM = 700;
const int DISPLAY_MIN_Z = 300;
const int DISPLAY_MAX_Z = 3000;

• MAX_FOV_DEPTH_MM defines the maximum distance (70 cm) of the points we want to keep in the foreground.
• DISPLAY_MIN_Z and DISPLAY_MAX_Z are the depth range used to normalize and colorize the depth map.
• These constants keep the depth filtering behavior consistent across the program.

cv::Mat colorizeDepthMap(const cv::Mat& depth, const int display_min_z = DISPLAY_MIN_Z, const int display_max_z = DISPLAY_MAX_Z) {

    const double normalize_alpha = 255.0 / static_cast<double>(display_max_z - display_min_z);
    const double normalize_beta = static_cast<double>(-255.0 * display_min_z) /
        static_cast<double>(display_max_z - display_min_z);

    cv::Mat output, depth_normal;

    depth.convertTo(depth_normal, CV_8UC1, normalize_alpha, normalize_beta);

    cv::applyColorMap(depth_normal, output, cv::COLORMAP_JET);
    cv::Mat displayMask = (depth >= display_min_z) & (depth <= display_max_z);
    cv::Mat fovMask = (depth <= MAX_FOV_DEPTH_MM);
    cv::Mat finalMask = displayMask & fovMask;
    output.setTo(Scalar(0, 0, 0), ~finalMask);
    return output;
}

cv::Mat colorizeDepthMap(const cv::Mat& depth, const int display_min_z = DISPLAY_MIN_Z, const int display_max_z = DISPLAY_MAX_Z) {

    const double normalize_alpha = 255.0 / static_cast<double>(display_max_z - display_min_z);
    const double normalize_beta = static_cast<double>(-255.0 * display_min_z) /
        static_cast<double>(display_max_z - display_min_z);

    cv::Mat output, depth_normal;

    depth.convertTo(depth_normal, CV_8UC1, normalize_alpha, normalize_beta);

    cv::applyColorMap(depth_normal, output, cv::COLORMAP_JET);
    cv::Mat displayMask = (depth >= display_min_z) & (depth <= display_max_z);
    cv::Mat fovMask = (depth <= MAX_FOV_DEPTH_MM);
    cv::Mat finalMask = displayMask & fovMask;
    output.setTo(Scalar(0, 0, 0), ~finalMask);
    return output;
}

• Converts a 16-bit depth image into a colorized visualization using cv::applyColorMap.
• Normalizes depth values into 0–255 based on display_min_z and display_max_z.
• Creates a mask that keeps only pixels:
• Within the display range
• Within the FOV limit (MAX_FOV_DEPTH_MM, i.e. 70 cm).
• Pixels outside this combined mask are set to black – effectively “removing” background points beyond the desired depth.

 bool regEnabled = false;
    if (device.isImageRegistrationModeSupported(IMAGE_REGISTRATION_DEPTH_TO_COLOR))
    {
        if (device.setImageRegistrationMode(IMAGE_REGISTRATION_DEPTH_TO_COLOR) == STATUS_OK)
        {
            regEnabled = true;
            cout << "[INFO] Registration enabled: DEPTH -> COLOR\n";
        }

 bool regEnabled = false;
    if (device.isImageRegistrationModeSupported(IMAGE_REGISTRATION_DEPTH_TO_COLOR))
    {
        if (device.setImageRegistrationMode(IMAGE_REGISTRATION_DEPTH_TO_COLOR) == STATUS_OK)
        {
            regEnabled = true;
            cout << "[INFO] Registration enabled: DEPTH -> COLOR\n";
        }

• Registration aligns depth pixels to the color camera viewpoint, which is critical when depth is used to mask RGB images.

cv::Mat rgbDisplayMask = (depthRaw >= DISPLAY_MIN_Z) & (depthRaw <= DISPLAY_MAX_Z);
cv::Mat rgbFovMask = (depthRaw <= MAX_FOV_DEPTH_MM);
cv::Mat rgbFinalMask = rgbDisplayMask & rgbFovMask;

cv::Mat rgbDisplayMask = (depthRaw >= DISPLAY_MIN_Z) & (depthRaw <= DISPLAY_MAX_Z);
cv::Mat rgbFovMask = (depthRaw <= MAX_FOV_DEPTH_MM);
cv::Mat rgbFinalMask = rgbDisplayMask & rgbFovMask;

• Builds a display range mask that keeps only depth values between DISPLAY_MIN_Z and DISPLAY_MAX_Z.
• Builds a FOV mask that keeps only points closer than or equal to MAX_FOV_DEPTH_MM (70 cm).
• Combines both masks to get the final foreground region.
• Applies the inverse of this mask to the RGB image to set background pixels to black.
• Effectively simulates a depth-based foreground extraction, similar in concept to GrabCut but using depth instead of interactive user input.