State-of-the-art evaluation of 'Shape from Polarisation' methods

Abstract

Advancements in the fields of Computer Vision and Artificial Intelligence have resulted in the emergence of innovative techniques for the extraction of three-dimensional shape information. One such technique is Shape from Polarization (SfP), which tries to exploit polarization patterns of light to achieve its goal. This thesis provides a comprehensive analysis of the most recent state-of-the-art SfP techniques, with a particular focus on comparing three physics-based methodologies and a hybrid approach that fuses Deep Learning with physics-based modelling. The study aims to assess the strengths and limitations of each technique through a systematic assessment evaluation using a dataset of rendered images. To ensure fair comparison and evaluation, a tailored dataset of synthetically created polarization images was created using the Mitsuba 3 rendering engine. The comparative analysis utilizing the tailor-made dataset provides valuable insights into the performance of each technique. This serves as a useful guide for researchers and practitioners in selecting appropriate methods based on specific application requirements, while also providing insight into the usability or otherwise of these methods in real-world contexts.