What does ignoring tangential lens distortion cost a self-calibrating bundle adjustment?

In simulation, setting Brown tangential distortion to zero doubles the 3D error on a wide rig at p = 0.01 and corrupts the fitted k2 on a narrow one.

Camera calibration inside a bundle adjustment usually fits focal length and radial distortion. The Brown (1966) model also has two tangential terms, p1 and p2, for a lens that is slightly decentred. How much does leaving them out cost? I ran it on the synthetic object-orbit rig the sensing-* projects share (12 views, 0.5 px noise, 8 seeds, every starting pose wrong).

With p1 = 0.01 and p2 = -0.008 in the images and the solver holding both at zero, the surface error on the wide rig (f = 400 px) goes from 0.0018 to 0.0041 object radii, and to 0.0077 at p1 = 0.02. The narrow rig (f = 800 px) is half as sensitive in surface error. Its fitted k2 is the surprise: at p1 = 0.02 it comes out at +0.45 against a truth of +0.10, because k2 is barely identified at that range and the missing tangential term lands in it. Fitting p1, p2 recovers them (0.0086 and -0.0080 for 0.0100 and -0.0080) and holds the error flat. At 2 px of pixel noise the benefit disappears.

Caveats: synthetic only, p is swept rather than taken from a measured calibration, the principal point is held at the centre in both fits, so the principal-point/tangential confound is untested. That is the next step.

Code, tests and the results table: research/sensing-bapp.

All writing · More in Research findings · All tools and research