Date Available

3-19-2023

Year of Publication

2023

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Engineering

Department/School/Program

Computer Science

Advisor

Dr. Brent Harrison

Co-Director of Graduate Studies

Dr. Nathan Jacobs

Abstract

The appearance of scenes may change for many reasons, including the viewpoint, the time of day, the weather, and the seasons. Traditionally, deep neural networks are trained and evaluated using images from the same scene and domain to avoid the domain gap. Recent advances in domain adaptation have led to a new type of method that bridges such domain gaps and learns from multiple domains.

This dissertation proposes methods for multi-domain adaptation for various computer vision tasks, including image classification, depth estimation, and semantic segmentation. The first work focuses on semi-supervised domain adaptation. I address this semi-supervised setting and propose to use dynamic feature alignment to address both inter- and intra-domain discrepancy. The second work addresses the task of monocular depth estimation in the multi-domain setting. I propose to address this task with a unified approach that includes adversarial knowledge distillation and uncertainty-guided self-supervised reconstruction. The third work considers the problem of semantic segmentation for aerial imagery with diverse environments and viewing geometries. I present CrossSeg: a novel framework that learns a semantic segmentation network that can generalize well in a cross-scene setting with only a few labeled samples. I believe this line of work can be applicable to many domain adaptation scenarios and aerial applications.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2023.038

Download

Share

COinS