We propose a framework for modeling the action of occluder-based intermediate frames for the purpose of computational imaging, and describe an information-theoretic metric for evaluating these frames. We propose a simple method for choosing an occluder when optimal constructions are unknown or do not exist. We validate the performance of this method relative to other methods empirically.
We also describe accidental-camera NLoS-imaging methods for reconstructing hidden scenes under a variety of conditions, including a method for recovering 1D reconstructions of scenes around a corner, recovering light-fields using a calibrated system, reconstructing a moving scene with an unknown occluder, and recovering a moving scene from a heterogeneous observation.
Occluders, i.e. opaque objects, can be used in the apertures of cameras to supplement or replace a traditional lens. This thesis describes a novel mutual-information theoretic framework for analyzing and comparing occluders. It justifies the use of uniformly-redundant arrays (URAs), a popular choice of pattern in coded-aperture imaging, and compares a method for selecting between different URAs to other methods of choosing occluders, such as a greedy search, and finds each to be preferable under different conditions. It also shows, analytically and empirically, the superiority of designed occluding patterns like URAs to random occluding patterns. The mutual-information theoretic framework is compared to a similar, MSE-minimizing framework.
This thesis also considers the use of occluders in the context of non-line-of-sight (NLoS) imaging, used as ``accidental cameras.'' The idea of the accidental camera is to opportunistically make use of occluding objects that happen to be available as ad-hoc coded apertures. Methods of this class, having originally been developed by Torralba and Freeman in 2012, are extended in this thesis to a wide variety of different scenarios. These include imaging around a corner using the corner as the occluder, imaging a light-field of an unknown scene using a known, calibrated occluder, imaging an unknown scene using an unknown occluder, and imaging an unknown scene with a heterogeneous observation using the deep image prior for blind matrix factorization. This thesis also uses the the tools of the earlier-mentioned framework for NLoS imaging, both for scene reconstruction, and to draw tentative conclusions about the maximum wavelength of light that can be used for NLoS imaging which will not bottleneck reconstruction quality.
The contributions of this thesis illustrate the value of occluder-based imaging and the powerful and flexible framework used to analyze it.
Thesis Supervisor: Prof. Greg Wornell
To attend this defense, please contact the doctoral candidate - adamyedidia at gmail dot com