Electronic Theses and Dissertations

Date

2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Electrical & Computer Engineering

Committee Chair

Madhusudhanan Balasubramanian

Committee Member

Deepak Venugopa

Committee Member

Eddie Jacobs

Committee Member

Mohammadreza Davoodi

Abstract

This dissertation presents a comprehensive framework for learning accurate, and physically meaningful representation of motion from image sequences. The central theme of this work lies in advancing optical flow estimation from discrete-time two-frame formulation toward a continuous time multi-frame model capable of learning higher-order motion terms and dynamics, with wide ranging applications including developing structural disease biomarkers from medical images. We first develop SSTM (Spatiotemporal Recurrent Transformer for Multi-Frame Optical Flow Estimation), a deep learning model that leverages a sequence of multi-frames to infer accurate optical flow by learning temporal dependencies across successive motion cues. The architecture integrates recurrent 3D convolutional GRU blocks and spatiotemporal transformers to jointly model spatial and temporal correlations of moving elements. Through extensive experiments on standard benchmarks for optical flow such as Sintel and KITTI2015, SSTM achieved state-of-the-art accuracy compared to other multi-frame methods (as of 2023), and outperformed two-frame methods in highly occluded and out-of-boundary regions, showcasing the benefits of learning wider temporal cues for accurate optical flow estimation. Building on these formulations, we introduce CTFlow, a framework for continuous-time optical flow estimation trained entirely on image sequences. Unlike prior methods that exploit high temporal resolution of event cameras, trained and validated on standard optical flow benchmarks, Sintel and KITTI2015, CTFlow models motion as dense, smooth, and differentiable trajectories parameterized by B´ezier curves. The ability of CTFlow to model higher order motion terms, such as instantaneous velocity and acceleration from sparse frame sequences, paves a way for advancements in other related computer vision tasks, such as video frame interpolation, extrapolation and compression. Finally, we translate these optical flow advances into a clinical application for the early detection of glaucoma progression. We introduce novel optical flow-derived structural biomarkers of glaucoma from longitudinal optic nerve head (ONH) images. These kinematic descriptors include strain descriptors and vorticity descriptors which are capable of capturing local deformation patterns in shape associated with tissue expansion and contraction as well as local rotations and spins, respectively, those jointly defining the underlying deformation during progressive glaucoma. A major challenge in modeling structural changes associated with biological phenomena such as growth, disease and aging is the absence of ground-truth data or observations. This is because the intrinsic forces driving these changes are too complex and difficult to directly measure or model. To mitigate this, we developed synthetic and naturalistic dataset generation algorithms and pipelines for fine-tuning our models. The standard optical flow benchmarks often model rigid and multi-object motion with scene elements elements moving or reorganizing independently without any interacting forces acting across their boundaries. Therefore, for modeling biological remodeling, we synthesize datasets that emulate single body deformation where various forces act a single point to enact tissue and organ reorganization. Models fine-tuned using these synthetic sequences were validated using experimental tensile test sequences of elastic specimen with known mechanical properties and tested on clinical datasets of optic nerve head changes in glaucoma.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest/Clarivate.”

Notes

Open Access

Download

Share

COinS
 

Archival Statement

This item was created or digitized prior to April 24, 2026, or is a reproduction of legacy media created before that date. It is preserved in its original, unmodified state specifically for research, reference, or historical recordkeeping. This material is part of a digital archival collection and is not utilized for current University instruction, programs, or active public communication. In accordance with the ADA Title II Final Rule, the University Libraries provides accessible versions of archival materials upon request. To request an accommodation for this item, please submit an accessibility request form.