after reading a few more papers, the amazing thing is they are using this technology to restore point clouds of landmarks off random tourist photos. bloody amazing.
this is what i read you got to do... and it aint easy.
ok, so first off, youve got to find the same 3d locations on all the photos
this can actually be done manually, if you have the time.
doing this automatically sounds nearly impossible.
you have to reconstruct the 3d positions orientations and fov from this alone. what the fuck!??
sounds like really powerful algebra at work... even with multiple valid solutions.
then after that, the triangulation problem is errored to shithouse.
the more picky you are with validating points, the better the result, but the longer it takes.