Question

I am using openCV’s cvMatchTemplate method to find a smaller image in a bigger image. The \ method below is matching the existing pattern object with the given ‘current’ image, the found pattern is marked by a rectangle drawn on the current image. When the pattern is visible its finding it reliably, when not its drawing somewhere “random” which makes sense.

I have trouble computing the score of the template matching. Is there an easy to determine that? I was looking on various sites but could not find a solution and also been looking at the values computed by cvMinMaxLoc but could not find a way to determine the quality of the match.

- (void)detect:(IplImage *)current
{
int patchx = pattern->width;
int patchy = pattern->height;
int iwidth = current->width - patchx + 1;
int iheight = current->height - patchy + 1;

IplImage *result=cvCreateImage( cvSize(iwidth,iheight),IPL_DEPTH_32F, 1);
cvMatchTemplate( current, pattern, result, CV_TM_SQDIFF     );

CvPoint     minloc, maxloc;
double      minval, maxval;

cvMinMaxLoc( result, &minval, &maxval, &minloc, &maxloc, 0 );

/* draw red rectangle */
cvRectangle( current, 
            cvPoint( minloc.x, minloc.y ), 
            cvPoint( minloc.x + patchx, minloc.y + patchy ),
            cvScalar( 0, 255, 0, 0 ), 1, 0, 0 );
}

Answer 1

Use a normed metric like CV_TM_CCORR_NORMED or CV_TM_SQDIFF_NORMED and compare the minimum value to a threshold (between 0 and 1.0).

cvMatchTemplate( current, pattern, result, CV_TM_SQDIFF_NORMED);

if (minval < THRESHOLD) {
    /* draw red rectangle */
    cvRectangle( current, 
            cvPoint( minloc.x, minloc.y ), 
            cvPoint( minloc.x + patchx, minloc.y + patchy ),
            cvScalar( 0, 255, 0, 0 ), 1, 0, 0 );
} else {
    //not found
}