ORCID

0000-0002-0311-6135

Keywords

Human Pose Estimation, 3D Computer Vision, Transfer Learning, Domain Adaptation, Domain Generation, Generative AI

Abstract

This dissertation advances unsupervised transfer learning by addressing critical constraints in data privacy, intellectual property protection, and deployment in unseen environments. We propose novel frameworks for three progressively restrictive settings: Black-Box Domain Adaptation (BBDA), Source-Free Domain Adaptation (SFDA), Domain Generalization (DG), and Lifelong Domain Adaptation (LLDA) for 3D human pose estimation (HPE). To address BBDA, where only model outputs are accessible, we introduce RAIN (Regularization on Input and Network). This framework utilizes Phase MixUp for frequency-domain input regularization and Subnetwork Distillation to combat source overfitting.

Extending these privacy-preserving paradigms to spatial tasks, we present the first SFDA framework for HPE. By projecting heatmaps into a novel spatial probability space, we reduce dimensionality by 32× while preserving complete structural information. This approach explores the problem from both source-protect and target-relevant perspectives, outperforming existing methods by over 3.0% across RHD and SURREAL adaptation tasks. Furthermore, we tackle DG for 3D HPE by proposing a Dual-Augmentor Framework (DAF). Unlike single-augmentor methods, DAF employs weak and strong pose augmentors with a two-stage discrimination process and metaoptimization to simulate domain shifts, resulting in significant MPJPE reductions.

Finally, the dissertation introduces the first framework for Lifelong Domain Adaptive 3D HPE to handle non-stationary environments. We propose a unified 3D pose generator featuring a triple encoding strategy—comprising pose-aware, temporal-aware, and diffusion-based domain-aware encoders—to capture domain-specific priors while mitigating catastrophic forgetting. These contributions provide principled solutions for spatial sparsity, sequential data shifts, and privacy constraints, bringing computer vision systems closer to robust real-world deployment.

Completion Date

2026

Semester

Spring

Committee Chair

Chen Chen

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Format

PDF

Document Type

Dissertation

Identifier

DP0053111

Download

Share

COinS
 

Accessibility Statement

This item was created or digitized prior to April 24, 2027, 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. 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.