Ask Your Question

chamfer Matching error in implementation

asked 2017-06-20 04:28:34 -0500

Franz Kaiser gravatar image

updated 2018-10-05 00:07:06 -0500

I found a implimentation of chamfer Matching here , which seems to have an error in this line:

Point& new_point = Point(model_column,model_row);

-> see beraks comment - thank you!

The program runs, but the results are not as i expected. I translated the image 7 Pixels in each direction and still get (0,0) as a match, because the matching image is just 1x1 px.

I would divide the matching part in the following steps: 1. model points from canny output are stored in a vector 2. A matching space is created ->*if the model dimensions are substracted, does this mean, that the template has to fit on the image ? * 3. For every templatepoint the value of distance transform is added to a matching score. This is where i especially don't understand the following line:

        matching_score += (float) *(chamfer_image.ptr<float>(model_points[point_count].y+search_row) +
        search_column + model_points[point_count].x*image_channels);

Thank you for your help:

whole code:

cv::Mat imgTranslate(cv::Mat src, int col,  int dx, int dy)
    cv::Mat dst(src.size(), src.type(), cv::Scalar::all(col) );
    src(cv::Rect(dy,dx, src.cols-dy,src.rows-dx)).copyTo(dst(cv::Rect(0,0,src.cols-dy,src.rows-dx)));
    return dst;

void ChamferMatching( Mat& chamfer_image, Mat& model, Mat& matching_image )
    // Extract the model points (as they are sparse).
    vector<Point> model_points;
    int image_channels = model.channels();
    for (int model_row=0; (model_row < model.rows); model_row++)
        uchar *curr_point = model.ptr<uchar>(model_row);
        for (int model_column=0; (model_column < model.cols); model_column++)
            if (*curr_point > 0)
                const Point& new_point = Point(model_column,model_row);
            curr_point += image_channels;
    int num_model_points = model_points.size();
    image_channels = chamfer_image.channels();
    // Try the model in every possible position
    matching_image = Mat(chamfer_image.rows-model.rows+1, chamfer_image.cols-model.cols+1, CV_32FC1);
    for (int search_row=0; (search_row <= chamfer_image.rows-model.rows); search_row++)
        float *output_point = matching_image.ptr<float>(search_row);
        for (int search_column=0; (search_column <= chamfer_image.cols-model.cols); search_column++)
            float matching_score = 0.0;
            for (int point_count=0; (point_count < num_model_points); point_count++)
                matching_score += (float) *(chamfer_image.ptr<float>(model_points[point_count].y+search_row) + 
                search_column + model_points[point_count].x*image_channels);
               *output_point = matching_score;

int main()
    Mat templateImage = imread(img1, IMREAD_GRAYSCALE);
    Mat queryImage = imgTranslate(templateImage, 255, 7, 7);

    Mat edge_image, chamfer_image, model_edge;
    Canny( queryImage, edge_image, 100, 200, 3);
    threshold( edge_image, edge_image, 127, 255, THRESH_BINARY_INV );
    distanceTransform( edge_image, chamfer_image, CV_DIST_L2, 3);

    Canny( templateImage, model_edge, 100, 200, 3);

    Mat resultImage;
    ChamferMatching(chamfer_image, model_edge, resultImage);

    double min, max;
    cv::Point min_loc, max_loc;
    cv::minMaxLoc(resultImage, &min, &max, &min_loc, &max_loc);

    cout << min_loc << endl;
    return 0;
edit retag flag offensive close merge delete


"but the results are not as i expected." -- this might need more detail ;)

taking a reference of a locally constructed Point is kinda pointless, however, the correct expected thing would be:

const Point& new_point = Point(model_column,model_row);

(VS compilers arent that picky here)

still, more info on the results, please.. !

berak gravatar imageberak ( 2017-06-20 05:40:00 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2017-06-21 01:09:59 -0500

berak gravatar image

updated 2017-06-21 01:11:32 -0500

wait, this assumes to find a smaller template (the model) in a larger query image (chamfer_image)

you seem to have that in reverse . if you have only 1 image, the model should be a cropped roi of your query, not the other way round.

edit flag offensive delete link more


Thanks again, berak. As you said, my problem with the code above was the way the results are stored in a Mat: First entry is the match between left upper corner from template with left upper corner from queryimage. The question can be closed now.

Franz Kaiser gravatar imageFranz Kaiser ( 2017-06-21 11:12:06 -0500 )edit

Question Tools

1 follower


Asked: 2017-06-20 04:28:34 -0500

Seen: 333 times

Last updated: Jun 20 '17