Polarization in Computer Vision & Graphics

Abstract

Incorporating polarization in computer graphics and vision algorithms has been a growing area of interest due to the rise in the availability of commercial polarimetric cameras and the ability to track polarization in physics-based rendering (PBR) engines. This dissertation contributes to the areas of computer vision and graphics, with an emphasis on polarimetric applications. In computer vision, an area of interest is developing methods and techniques to optimize the performance of convolutional neural networks (CNN) in a binary image classification task with limited amounts of training data. The techniques discussed in this dissertation introduce a linear processing method that compresses and transforms image data in a way that can increase CNN detection performance. In addition to improving the detection of CNNs for binary classification, this dissertation demonstrates methods to increase CNN performance for road scene object detection. Here, polarimetry is used as an additional image modality to an existing CNN architecture, resulting in an increase in the detection performance of objects with low contrast, without the need to retrain the network. The work in this dissertation addresses the growing interest in implementing and analyzing polarized computer graphics rendering algorithms. Here, a new method for compressing, interpolating, and sampling a database of polarized bidirectional reflection distribution functions (pBRDFs) is introduced, which retains the dominant polarimetric process when rendered and ensures physicality.