[lazide]

Also, a typical consumer camera is going to struggle (even with the best analysis!) to have visual accuracy within even a few degrees of true (due to everything from high compression ratios in post processing to shitty optical properties in the actual camera) and that makes a probability cone that is pretty large.

Still better than nothing, though!

[anovikov]

building a WCS model (pixel XY <-> ray vectors) of a camera is a solved science. Of course you can never expect it to be rectangular.

[lazide]

No kidding, that is what I was referring too. Precision for a typical doorbell camera is going to be several minutes of angle at best, which is my point.

[anovikov]

Oh well several arcminutes is excellent. Yes this is best case.