As Saijin said above, it can be applied to GPS tagged images! (I believe it may be a requirement for Pix4D scale-constraints even).
I’ll have to go read some white papers on the exact math behind it, but the user-facing process starts after processing your initial orthomosaic. You find a point-of-interest with known dimensions and input coordinate points with the known lengths in-between them (it can also be used as an additional check when using well surveyed GCPs by checking the distance between them). So basically the program is saying this point_A is at (x,y) and point_B is at (x,y). The user then inputs that the distance between point_A and point_B is (n), the software mulls this over and makes some scale corrections by partially reprocessing the dataset.
You could use something as functionally specific as a yard stick that you’ve laid out, or just something with known measurements like a house or a car.
My experience with Pix4D and Agisoft is that scale-constraints do have reasonable limits of functionality (as they should). So say you’ve made a map with only GPS-tagged images and you create a scale-constraint with a previously placed yard stick. If the software originally calculated the yard stick as being 0.8 yards without scale-constraint corrections everything will most likely be able to scale appropriately. But if it had previously measured the yard stick as 0.1 yards, you’re going run into some issues. My point is that scale-constraints are an additional tool for relative accuracy correction, but they certainly don’t make significant changes possible and can’t correct poorly gathered data-sets.
Excuse my rambling, typing this up on a break and didn’t proof-read very well!