Cross-Domain Image Classification in Complex Real-World Scenarios: With Test-Time Label or Continual Domain Shift

Ma, Tianyi

Cross-Domain Image Classification in Complex Real-World Scenarios: With Test-Time Label or Continual Domain Shift

Ma, Tianyi

Permalink

Publication Type:: Thesis
Issue Date:: 2025

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download thesisAdobe PDF (3.51 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Ma, Tianyi
dc.date.accessioned	2026-06-30T02:54:51Z
dc.date.available	2026-06-30T02:54:51Z
dc.date.issued	2025
dc.identifier.uri	http://hdl.handle.net/10453/195484
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	Cross-domain image classification is proposed to address the distribution discrepancy between the source and target domains, commonly referred to as the domain gap. Although it has been proven effective in mitigating the domain gap between the source and target domains, cross-domain image classification still faces significant challenges in complex real-world scenarios. In this thesis, two specific challenges that commonly happen, namely, label shift and continual domain shift, for cross-domain image classification are investigated and addressed. For the label shift, this thesis focuses on cross-domain few-shot image classification (CDFSIC). A novel prompt-to-disentangle method is proposed to combine the benefits of domain generalisation and adaptation by disentangling source and target knowledge. For the continual domain shift, this thesis focuses on continual test-time adaptation (CoTTA). Based on the findings in CDFSIC, we design a similar strategy called the Source and Target Disentangle Transformer to explicitly disentangle source and target knowledge, thereby facilitating both the preservation of source knowledge and the extraction of target knowledge. Then, based on the observation that the recent CoTTA method is unstable in a small-batch setting, a novel task named single-sample CoTTA is proposed. A novel strategy, named effective buffer and resetting, is designed to increase adaptation stability. Moreover, we apply this method to zero-shot models, solving the label shift and continual domain shift simultaneously. Finally, we highlighted several future directions, including active CoTTA to address larger domain gaps with human calibrations and zero-shot CoTTA to tackle both label shift and continual domain shift.	en_US.UTF-8
dc.format	Thesis (PhD)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/195484/1/thesis.pdf
dc.rights	info:eu-repo/semantics/openAccess
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	© 2025 Tianyi Ma
dc.rights	au.edu.uts.lib/cph
dc.title	Cross-Domain Image Classification in Complex Real-World Scenarios: With Test-Time Label or Continual Domain Shift	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

Cross-domain image classification is proposed to address the distribution discrepancy between the source and target domains, commonly referred to as the domain gap. Although it has been proven effective in mitigating the domain gap between the source and target domains, cross-domain image classification still faces significant challenges in complex real-world scenarios. In this thesis, two specific challenges that commonly happen, namely, label shift and continual domain shift, for cross-domain image classification are investigated and addressed. For the label shift, this thesis focuses on cross-domain few-shot image classification (CDFSIC). A novel prompt-to-disentangle method is proposed to combine the benefits of domain generalisation and adaptation by disentangling source and target knowledge. For the continual domain shift, this thesis focuses on continual test-time adaptation (CoTTA). Based on the findings in CDFSIC, we design a similar strategy called the Source and Target Disentangle Transformer to explicitly disentangle source and target knowledge, thereby facilitating both the preservation of source knowledge and the extraction of target knowledge. Then, based on the observation that the recent CoTTA method is unstable in a small-batch setting, a novel task named single-sample CoTTA is proposed. A novel strategy, named effective buffer and resetting, is designed to increase adaptation stability. Moreover, we apply this method to zero-shot models, solving the label shift and continual domain shift simultaneously. Finally, we highlighted several future directions, including active CoTTA to address larger domain gaps with human calibrations and zero-shot CoTTA to tackle both label shift and continual domain shift.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/195484