Measuring Light Polarization via PFA Cameras: error modeling and analysis

Light is made of oscillating electric and magnetic fields, the orientation of these oscillations is called polarization. Since light polarization changes when it is reflected on a surface, polarization can be helpful in many computer vision tasks, including shape reconstruction, flaw detection, or material recognition, just to name a few. With the advent of Polarimetric Filter Array (PFA) cameras, academic interest in this area has significantly increased. PFA cameras are cameras equipped with linear polarizer filters in front of each pixel to measure light intensity at various polarization orientations, allowing them to measure the polarization state of light in a single shot. This thesis discusses the main weaknesses inherently present in PFA cameras, their impact and the possible mitigations. We will also present the state of the art of the Shape-from-Polarization techniques and analyze the existing polarimetric datasets. Furthermore, we propose a novel approach to model and minimize the error of PFA cameras acquisitions: such model provides the likelihood of a specific polarization state given the observed channels of a pixel and an estimation of the camera noise as inputs.