Deep Domain Adaptation for Visual Recognition
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Zhang, WeichenAbstract
Deep neural networks, which usually require a large amount of labelled data during training process, have achieved significant progress in many practical applications in this information era. Besides, due to the different environments and the variation of sensors used for data ...
See moreDeep neural networks, which usually require a large amount of labelled data during training process, have achieved significant progress in many practical applications in this information era. Besides, due to the different environments and the variation of sensors used for data collection, it often brings difficulties for general supervised deep learning methods to perform well in various scenarios. Therefore, deep transfer learning methods, especially the deep domain adaptation methods are developed to exploit the domain-invariant intrinsic features and improve the generalization capability of the deep neural networks. This thesis addresses several issues of domain adaptation for multiple practical applications. The main contributions of this thesis are listed below. First, a comprehensive literature review for deep learning, transfer learning, domain adaptation and multiple visual applications is presented to provide better understanding of the background of this thesis. Second, a general domain adaptation method called Collaborative and Adversarial Network (CAN) is introduced to jointly learn domain-specific and domain-invariant features through domain-collaborative and domain-adversarial training of deep neural networks. Third, a multi-modality domain adaptation framework called Progressive Modality Cooperation (PMC) is proposed for domain adaptation with multi-modality input, and is additionally extended to PMC using Privileged Information (PMC-PI) to deal with the domain adaptation problem when some modalities are missing in the target domain. Fourth, the Scale-aware and Range-aware Domain Adaptation Network (SRDAN) is designed to deal with the Cross-dataset 3D Object Detection task, where objects with similar geometric characteristics are guided to be aligned between two domains. Last, we give future research directions of domain adaptation.
See less
See moreDeep neural networks, which usually require a large amount of labelled data during training process, have achieved significant progress in many practical applications in this information era. Besides, due to the different environments and the variation of sensors used for data collection, it often brings difficulties for general supervised deep learning methods to perform well in various scenarios. Therefore, deep transfer learning methods, especially the deep domain adaptation methods are developed to exploit the domain-invariant intrinsic features and improve the generalization capability of the deep neural networks. This thesis addresses several issues of domain adaptation for multiple practical applications. The main contributions of this thesis are listed below. First, a comprehensive literature review for deep learning, transfer learning, domain adaptation and multiple visual applications is presented to provide better understanding of the background of this thesis. Second, a general domain adaptation method called Collaborative and Adversarial Network (CAN) is introduced to jointly learn domain-specific and domain-invariant features through domain-collaborative and domain-adversarial training of deep neural networks. Third, a multi-modality domain adaptation framework called Progressive Modality Cooperation (PMC) is proposed for domain adaptation with multi-modality input, and is additionally extended to PMC using Privileged Information (PMC-PI) to deal with the domain adaptation problem when some modalities are missing in the target domain. Fourth, the Scale-aware and Range-aware Domain Adaptation Network (SRDAN) is designed to deal with the Cross-dataset 3D Object Detection task, where objects with similar geometric characteristics are guided to be aligned between two domains. Last, we give future research directions of domain adaptation.
See less
Date
2021Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Electrical and Information EngineeringAwarding institution
The University of SydneyShare