Non-line-of-sight imaging using data-driven approaches

Abstract

Non-line-of-sight (NLOS) imaging is desirable for its many potential applications such as detecting a vehicle occluded by a building's corner or imaging through fog. Traditional NLOS imaging techniques solve an inverse problem and are limited by computational complexity and forward model accuracy. This thesis proposes the application of data-driven techniques to NLOS imaging to leverage the convolutional neural network's ability to learn invariants to scene variations. We demonstrate the classification of an object hidden behind a scattering media along with the localization and classification of an object occluded by a corner. In addition we demonstrate the use of generative neural networks to construct images from viewpoints that extend the original camera's field of view.