Hong Kong Baptist University (HKBU) Research Cluster on Data Analytics and Artificial Intelligence in X

Unsupervised Domain Adaptation without Source Data and Its Application to Visual Recognition
Principal Investigatgor: Prof. P.C. YUEN (Department of Computer Science)

Unsupervised domain adaptation has becoming a popular research area and many methods with promising results have been proposed in past few years. It is normally assumed that labelled (training) data in source domain is available. However, owing to the data privacy or commercial data confidentiality, company may not be allowed (or not willing) to distribute/share personal data to other parties without individual’s consent. For example, a new General Data Protection Regulation (GDPR) has been approved in April 2016 in Europe (to replace the existing data protection directive) and to be implemented in May 2018. GDPR has strengthen the restriction in transferring personal data from one electronic system to another. Moreover, under the GDPR, individual has the right to erasure his/her personal data from the system. As such, training data in source domain will not be available. Under this scenario, source data distribution is hard, if not impossible to estimate. With only a source classifier and target unlabled data, this project will propose to extract source and target regions from label neighbours which could (i) induce source and target data distributions, and (ii) facilitate cross domain alignment.


Objectives:

  1. To study and develop a new unsupervised domain adaptation framework in which only a source classifier and a large number of target unlabelled data, high-level statistical data in source domain are available. No raw source a is required; and
  2. To investigate and develop a new algorithm to extract useful information which could induce source and target data distribution information using only a source classifier and target unlabelled data; and
  3. To investigate and develop a new cross-domain alignment algorithm based on the information extracted in (2). After performing the information alignment, discriminative and adaptable target classifier can then be trained.


Related Publications:

  1. Jiawei Li, Andy Jinhua Ma, Pong C. Yuen, “Semi-supervised Region Metric Learning for Person Re-identification”, International Journal of Computer Vision, 126(8): 855-874, 2018.
  2. Rui Shao, Xiangyuan Lan, Pong C. Yuen, “Feature Constrained by Pixel: Hierarchical Adversarial Deep Domain Adaptation”, ACM Multimedia Conference (ACMMM), 2018.
  3. Baoyao Yang, Andy Jinhua Ma, Pong C. Yuen, “Domain-Shared Group-Sparse Dictionary Learning for Unsupervised Domain Adaptation”. AAAI, 2018.
  4. 4. Andy Jinhua Ma, Pong C. Yuen, Jiawei Li, “Domain Transfer Support Vector Ranking for Person Re-identification without Target Camera Label Information”, ICCV 2013.


Grant Support:

This project is supported by the Research Grants Council (RGC), Hong Kong SAR, China (Project HKBU12200518)



For further information on this research topic, please contact Prof. P.C. YUEN.