[DaBackpack]

Originally Posted by Matthia

Whenever something is brought up that involves making comparisons or contrasts among a certain population, I tend to take quantum levels into consideration. That being said, at a microscopic level, everything is different (iow, each item of the same type, like apples for instance, has a slightly different end-result structure of how it grew and is thus unique). Same applies to fingerprints, although it's extremely obvious that "nothing that has happened before will happen again with 100% accuracy and recreation".

I feel these kinds of questions need a definite variable of measurement when trying to settle down to a decision and come up with an answer. Like in terms of size, lengths, scaling, warping and etc., how closely related do these fingerprints need to be in order to consider them as "matching"?

When it comes to fingerprints, forensic analysts have a list of common features and patterns that they use to describe fingerprints. You might see something like this in crime shows where a digital fingerprint image has little green circles around some of the curves and edges. This is because you can't rely on direct image matching between two fingerprints: the same person won't always press their fingers down in the same exact angle or orientation every single time. To match a fingerprint, you need to match as many of these key features as possible.

In image recognition tasks there is a task called "alignment", wherein the program has to determine if two different images are showing the same object or scene from different POVs or if they are from two different objects/scenes. It's pretty difficult but there are ways to make the problem simpler by creating simpler representations of the images in question--- i.e. not using full RGBA image rasters. For us, when it comes to fingerprints, these "simpler representations" manifest as a set of "arches", "whorls", and "loops." After first observing prominent examples of these features in each fingerprint, you can try to construct a feature mapping from print A onto print B. (As an analogy, think about rotating one Rubix cube until it looks like another Rubix cube.) If such a mapping exists with a certain confidence threshold, you might decide that both prints came from the same finger. Otherwise, they could be from different people, or you might need better prints.

Fingerprinting unfortunately has not been as historically reliable as you might think, because the theoretically foundations were not well understood, and humans are prone to error when manually annotating prints with feature labels. The practice seems to have become way more reliable over the past decade or so with the mass adoption of convolutional neural networks and other data-driven computational approaches, and now fingerprinting is a pretty common way to verify identity on mobile devices. (I still don't trust the technology enough to bet my bank account on it though, lol)

To answer the OP: the answer is basically yes. But remember that analysts don't use prints from a single finger: we have 10 fingers with unique data on them. Even if two people's index fingers might look very similar, you will see clear differences when looking at the 9 other fingers.