Graduation Year

2026

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

John Murray-Bruce, Ph.D.

Committee Member

Shaun Canavan, Ph.D.

Committee Member

Lawrence Hall, Ph.D.

Committee Member

Kaiqi Xiong, Ph.D.

Committee Member

Alfredo Weitzenfeld, Ph.D.

Committee Member

Ashwin B. Parthasarathy, Ph.D.

Keywords

computational imaging, deep learning, diffusion models, generative models, inverse problems, non-line-of-sight imaging, vision-language model, large-language model, 3D reconstruction, physics-based modeling, multimodal learning, occluder-aided, implicit neural networks

Abstract

Non-line-of-sight (NLOS) imaging, or seeing around corners, is the ability to recover information about objects hidden from direct view and remains one of the most compelling challenges in computational imaging. Existing approaches fall broadly into two categories: active and passive. Active time-resolved methods achieve impressive performance but require specialized pulsed laser hardware and ultrafast detectors, making them expensive and limited by slow measurement acquisition. Passive NLOS methods, which instead exploit the subtle soft shadows (penumbrae) cast by a hidden scene onto a visible matte surface, have recently emerged as a practical alternative. However, existing passive approaches are largely limited to one-dimensional video reconstruction, low-resolution two-dimensional color imaging, or localization of objects whose shape is assumed to be known. To enable practical real-world deployment of passive NLOS imaging systems, it is therefore essential to relax these restrictive assumptions and develop methods capable of recovering higher-dimensional scene structure under minimal prior knowledge.

This dissertation develops a comprehensive suite of passive NLOS imaging methods that address this challenge. We first reformulate the problem as the joint recovery of two unknown quantities: the three-dimensional geometry of light-occluding structures within the hidden scene and the two-dimensional full-color radiosity of the surrounding hidden environment. We then develop methods that progressively recover both quantities from a single photograph captured by an ordinary camera under ambient illumination. The unifying technical foundation is a novel reformulation of the light transport model that leverages pinspeck properties to decompose the hidden scene into light-occluding components---represented using multiple geometric models, including coarse three-dimensional binary occupancy grids, point clouds, and signed distance functions---and non-occluding components, represented as a transverse two-dimensional radiosity map. This decomposition transforms the reconstruction task into a separable nonlinear least-squares inverse problem.

Building on this formulation, the four contributions of this dissertation advance passive NLOS imaging through a deliberate progression: from physics-driven reconstruction, to hybrid physics--generative modeling, to fully learned end-to-end reconstruction, and finally to foundation-model-driven semantic inference, with each contribution addressing the limitations of its predecessor. Extensive validation on both synthetic and real-world experimental data demonstrates that these methods collectively advance passive NLOS imaging from proof-of-concept to an experimentally viable technology. Taken together, these advances provide a unified perspective on how physical models, learned priors, and foundation models complement one another in solving severely ill-posed inverse problems. Collectively, these contributions establish a unified framework that advances passive NLOS imaging in both computational reconstruction and high-level scene understanding, positioning passive NLOS imaging as a practical method for seeing around corners.

Download

Share

COinS