initialize_cameras_landmarks should support per-frame initial intrinsics about kwiver

[mwoehlke]

Ah... is sfm_constraints intended for this? Is it safe, then, to set the principal point to any old value? (At least the documentation has an error; it mentions a metadata parameter, but the parameter is actually constraints.)

[mleotta]

The principal point is almost always assumed to be the center of the image and fixed at the center. Estimating a custom principal point for imagery taken from a normal camera is usually overfitting the model unless you have a specialty camera where you know the lens is not centered with the sensor.

That aside, in general it would be good to be able to specify an initial guess for camera intrinsics on each camera. sfm_constraints is more for optimization constraints (e.g. make sure this camera stays close to this GPS location). For initialization I suggest we put the initial guess in the input camera map. We could add a config option that tells the algorithm to only use the intrinsics from those initial guesses.

[mwoehlke]

Riiiiight... although for my usage I have the unenviable case that my inputs may have been (perhaps even probably have been) cropped, which means the principle point is I-have-no-idea-where. I'd hope I can ignore that, but... that's actually a good question; can I? And, more to the point, if my inputs are of wildly different sizes (say, 200×200 and 2000×2000), am I going to run into trouble not being able to specify a separate principle point for each of those?

It sounds like constraints is not the mechanism for dealing with this?

For initialization I suggest we put the initial guess in the input camera map.

From my (GUI) end this should be pretty straight-forward; in fact, I'd even started writing it that way before realizing this was an issue. I guess this would need changes in KWIVER though? (Hence why I opened the issue...)