Mitigating Domain Shifts in Real-world Image Recognition

The School of EECS is hosting the following seminar:

Mitigating Domain Shifts in Real-world Image Recognition

Speaker: Zixin Wang
Host: Professor Xue Li

Abstract: Real-world image recognition models often suffer significant performance declines when there is a mismatch between the training (source) and deployment (target) domains, a phenomenon known as domain shifts. Domain adaptation (DA) has proven highly effective in addressing these shifts that the target data diverges from the training data due to non-i.i.d. conditions such as variations in style, location, lighting, weather conditions.

This thesis focuses on practical scenarios of domain adaptation. To adapt models to target domains that may contain new, unseen categories, we first introduce a variational domain disentanglement (VDD) framework. VDD separates domain-specific features from semantic content, enhancing the model's ability to generalize to shifted target sets. To address source domain privacy during adaptation, we propose Cal-SFDA, leveraging the expected calibration error to guide adaptation, ensures reliable predictions for real-world data. The final part of the thesis transitions to real-time adaptability with Online Test-Time Adaptation (OTTA). We explore how models can dynamically adjust to changes as new data streams in, a requirement for deployment in constantly evolving environments. Using modern architectures like Vision Transformers, the research benchmarks different OTTA methods, highlighting their strengths and limitations, and ensures that models remain effective despite ongoing shifts in input data distribution.

To conclude, this thesis provides a comprehensive study for domain adaptation, progressing from static, complex scenarios to dynamic, real-time adaptation. Together, they form a cohesive narrative that addresses the need for scalable, secure, and effective solutions, ensuring robust model performance in diverse and evolving real-world environments.

Bio: Zixin Wang is a final-year PhD student in the Data Science group at the School of EECS, supervised by Dr Sen Wang, Dr Yadan Luo, and Dr Zi Huang. Her research centres on domain adaptation for image recognition tasks. She received her master’s degree from The University of Queensland. Her work has been featured in conferences and journals such as ACM Multimedia, IJCV, etc, and received Best Student Paper Award at ACM Multimedia 2023.