Non-parametric Models of Distortion in Imaging Systems.

Abstract

Traditional radial lens distortion models are based on the physical construction of lenses. However, manufacturing defects and physical shock often cause the actual observed distortion to be different from what can be modeled by the physically motivated models.

In this work, we initially propose a Gaussian process radial distortion model as an alternative to the physically motivated models. The non-parametric nature of this model helps implicitly select the right model complexity, whereas for traditional distortion models one must perform explicit model selection to decide the right parametric complexity.

Next, we forego the radial distortion assumption and present a completely non-parametric, mathematically motivated distortion model based on locally-weighted homographies. The separation from an underlying physical model allows this model to capture arbitrary sources of distortion. We then apply this fully non-parametric distortion model to a zoom lens, where the distortion complexity can vary across zoom levels and the lens exhibits noticeable non-radial distortion.

Through our experiments and evaluation, we show that the proposed models are as accurate as the traditional parametric models at characterizing radial distortion while flexibly capturing non-radial distortion if present in the imaging system.