Non normalized matching functions are crucial to differentiate between objects with similar shape but different colors.
Here is a general example:
Suppose we look for a bright circle, a bright line or any other simple shape on a darker image. If we use a normalized distance, then most bright circles, lines, etc. (of similar size and shape) will get a good matching grade, regardless of their actual values. In non-artificial images, it is most likely that we will be flooded with a huge number of matches, which differs a lot from the template, but happen to have similar intensity variations. This flood of matches is often unfeasible to handle.
Here is a more specific example:
Lets assume that we track a ball in a basketball game. The ball color remains constant between frames, but if we normalize the match, many of the players and spectators' heads will get a good match, even though the non normalized template would have detected the correct ball, due to its accurate color.
Here is another specific use for non-normalized matching functions:
In stereo matching, or object tracking we wish to match an object in one frame with the same object in another frame based on its accurate color, and not get lots of matches from objects which have similar shape and color variations but with very different color. Therefore, non-normalized matching is required.
why you can't use non-normalized CrossCorr for such purposes? For SqrDistance you look for minimum, for CroosCorr you should look for maximum - so what is the difference for you? Both are done via 2D FFT and have the similar performance. From tha manual (ippiman.pdf - page 664):
"There are several ways to compute the measure of similarity between two images. One way is to compute the Euclidean distance, or sum of the squared distances (SSD), of an image and a template. The smaller is the value of SSD at a particular pixel, the more similarity exists between the template and the image in the neighborhood of that pixel.
The other similarity measure is the cross-correlation function: the higher is the cross-correlation at a particular pixel, the more similarity exists between the template and the image in the neighborhood of that pixel."
Non normalized cross correlation often gives a good match to a template and an image location even though the template is very different than the image. This happens when the image values are large. Here is a 1D example (in the 2D case the same phenomena described below occurs):
Given 1D Template values: 1, 1, 1,
Example a: 1D image values are: 2, 2, 2, 2 (result in small correlation with template).
Example b: 1D image values are: 1, 1, 1000, 1 (result in large correlation with template).
As you can see, the image in example (b) will get much larger correlation values even though it is very different from the template, while the image in example (a) which has values which are close to the template values, will get a lower correlation value than in example (b).
That is why unless the norm of both image and template are known, then one has to normalize the correlation.
This problem would not occur when using non-normalized square diff: The image in example (a) will result in much smaller square distance from the template than the image in example (b), as it should be.