Analyzing stereo-pairs of photographs is a matter of locating the same precise point as it is captured in two or more photographs. In the old days, this meant using a digitizing puck to magnify and locate the x,y location of a point (relative to a set of cross hairs that were essentially coplanar to the film). Today, with digital photographs, it is a matter of specifying pixels. In both cases, a handful of these comparison points is needed to construct a geometric model for each camera station: each point becomes the terminus of a theoretical “ray” that passes through the focal point of the lens. When the precise nature of how light passes through the focal point of the lens onto the film plane (or digital sensor) is understood mathematically, then this geometrical model can be quite precise, allowing for an accurate “back calculation” of the relative relationship of the camera from on photo station to another.
Its uncommonly difficult to write about – when you see it “happen” it is a little easier to understand!
Today, this whole process is becoming more and more automated, allowing for the creation of very rich point clouds when needed – or quickly creating smaller “smart point clouds” when the goal is a vector line drawing of a building, for example.