Wanna easy way to calculate the accuracy of template matching


This work began last week for a recognizing task, for which I have to calculate the accuracy of template matching and prefer sub-pixel accuracy if it is possible. The 1st thing I have to consider is the speed of existing approaches I can use in the project. The ones implemented via OpenCV were used for test and it seems acceptable while running on my laptop. Then, I come to the 2nd work, i.e. the accuracy.
Any suggestion would be appreciated.

maybe you can crop out various template rois from an image, and use euclidean distance between found location , and original cropped position

I don’t know about accuracy but resolution can certainly be gotten. you’d have to assess the accuracy from synthetic data (i.e. resampled/interpolated data where you know the shift).

use the response map from matchTemplate. if that doesn’t suit you, take a look at samples/python/mosse.py. it’s a correlation tracker.

whatever response map you have, it is possible to find the peak, take its neighborhood, assume a parabola or gaussian shape, and calculate the apex of that.