Question

Could anyone share their code for calculating the PGH (Pairwise Geometric Histogram) similarity ? I need to find the most similar object from within a list of images.

I have written up the following code, but the results make no sense. I’d bet I am doing a silly error, and I am stuck.

Any suggestions?

 public double GetBestPGHMatch(Contour<Point> currContour, List<Contour<Point>> ContoursList)
    {
        double match = -1.0d;
        DenseHistogram histCurrContour = new DenseHistogram(
                                                       new int[2]
                                                                {
                                                                    currContour.Total,
                                                                    currContour.Total,
                                                                },
                                                        new RangeF[2]
                                                                {
                                                                    new RangeF(0, 100),
                                                                    new RangeF(0, 100)
                                                                }
                                                   );

        CvInvoke.cvCalcPGH(currContour.Ptr, histCurrContour.Ptr);
        foreach (Contour<Point> contour in ContoursList)
        {
            DenseHistogram hist = new DenseHistogram(
                                              new int[2]
                                                                {
                                                                    currContour.Total,
                                                                    currContour.Total,
                                                                },
                                                new RangeF[2]
                                                                {
                                                                    new RangeF(0, 100),
                                                                    new RangeF(0, 100)
                                                                }
                                          );

            CvInvoke.cvCalcPGH(contour.Ptr, hist.Ptr);
            double c = CvInvoke.cvCompareHist(histCurrContour.Ptr, hist.Ptr, Emgu.CV.CvEnum.HISTOGRAM_COMP_METHOD.CV_COMP_CORREL);
            if (c > match) match = c;
        }

        return match;
    }

Answer 1

Hope this helps someone, here is how I worked it out, I am using Bhattacharya distance, hence the larger the value, the lower the match. There are other metrics as well, but I found B-distance best for my needs.

 public double pghMatchShape(Contour<Point> shape1, Contour<Point> shape2)
    {
        DenseHistogram hist1 = new DenseHistogram(
           new int[2] { 8, 8 },
           new RangeF[2] { new RangeF(-180, 180), new RangeF(100, 100) });
        DenseHistogram hist2 = new DenseHistogram(
           new int[2] { 8, 8 },
           new RangeF[2] { new RangeF(-180, 180), new RangeF(100, 100) });
        CvInvoke.cvCalcPGH(shape1, hist1.Ptr);
        CvInvoke.cvCalcPGH(shape2, hist2.Ptr);
        CvInvoke.cvNormalizeHist(hist1.Ptr, 100.0);
        CvInvoke.cvNormalizeHist(hist2.Ptr, 100.0);
        double corr = CvInvoke.cvCompareHist(hist1, hist2, HISTOGRAM_COMP_METHOD.CV_COMP_BHATTACHARYYA);

        return corr;
    }