Revision history [back]

to use the Mahalanobis distance, you need to build a global "model" first - the inverse covariance matrix of all your training samples. then, using that, you can compare a test sample to a train sample, to find out which is the best match in your database . so:

        //// step 1, offline / once:
        // put all your processed train samples into a single Mat, 
        //  each feature flattened onto a row of its own 
        Mat features;

        Mat _covar, _mean;
        calcCovarMatrix(features, _covar, _mean, CV_COVAR_NORMAL|CV_COVAR_ROWS, CV_32F);
        _covar /= (features.rows-1);

        Mat icovar; // this is the one used later
        invert(_covar, icovar, DECOMP_SVD);


        //// step 2, predict on a test image
        Mat testFeature;
        for (all trainFeatures) {
             Mat trainFeature = ...
             double dist = Mahalanobis(testFeature, trainFeature, icovar);
        }

so far, so good. however, be aware, that building a covariance matrix and even inverting it is extremely expensive, that's why most often some feature reduction like PCA is applied before.

to use the Mahalanobis distance, you need to build a global "model" first - the inverse covariance matrix of all your training samples. then, using that, you can compare a test sample to a train sample, to find out which is the best match in your database . so:

        //// step 1, offline / once:
        // put all your processed train samples into a single Mat, 
        //  each feature flattened onto a row of its own 
        Mat features;

        Mat _covar, _mean;
        calcCovarMatrix(features, _covar, _mean, CV_COVAR_NORMAL|CV_COVAR_ROWS, CV_32F);
        _covar /= (features.rows-1);

        Mat icovar; // this is the one used later
        invert(_covar, icovar, DECOMP_SVD);


        //// step 2, predict on a test image
        Mat testFeature;
        for (all trainFeatures) {
             Mat trainFeature = ...
             double dist = Mahalanobis(testFeature, trainFeature, icovar);
        }

so far, so good. however, be aware, that building a covariance matrix and even inverting it is extremely expensive, that's why most often some feature reduction like PCA is applied before.before (shorter features).