Unsupervised Domain Adaptation without Source Data and Its Application to Visual Recognition

Visual recognition has been an active research area for more than two decades. Some research results such as face recognition have been successfully applied in practical applications including mobile device logon, electronic payment systems and unmanned border control. One of the key success factors for transferring research results into practical application is that the proposed visual recognition methods have to be robust to different testing environments. However, in many visual recognition methods, such as person re-identification, the accuracy will be degraded when applying to a new environment under different conditions, e.g. change of sensors (cameras) and illumination. The level of degradation depends on the degree of distribution mismatch between data collected in the training and testing environments. It is known as data bias problem. To overcome the data bias problem, recent research study shows that a combination of a strong classifier (e.g. deep neural network) and domain adaptation (or transfer learning) is a promising direction. Existing domain adaptation normally assumes that (source domain) training data is available. However, this assumption will not be valid in many situations.

There are several reasons why training data in source domain will not be readily available. First, owing to data privacy or commercial data confidentiality, organisations/companies may not be willing (or allowed) to distribute or share personal data to other parties without consent. For example, the General Data Protection Regulation (GDPR) was approved in Europe in April 2016, and will be implemented in May 2018. The GDPR has strengthened the restrictions on transferring personal data from one electronic system to another. Moreover, under the GDPR, an individual has the right to have his or her personal data erased from a system. Second, existing strong classifier (e.g. deep neural network) requires a large amount (millions) of training data. Making such a large quantity of training data available to production applications is not practical.

Another issue in domain adaptation comes from the fact that labelling data (e.g. labelling video data for person re-identification) in target domain is time consuming and expensive. For that, the solution is to use unsupervised domain adaptation which avoids the need to label target domain data and thus has a higher practical value.

To handle above concerns, this project proposes to study and investigate a new and challenging scenario of unsupervised domain adaptation without source domain data.

The problem in unsupervised domain adaptation without source data that needs to be addressed is as follows. Given a well-trained source classifier and many unlabelled data in target domain, we would like to develop an adaptable discriminative target classifier on unseen target data. Detecting labels from target unlabelled data using source classifier is not a good approach to solve the problem, because source classifier is not reliable on target unseen data due to domain shift. Moreover, without source data, it is difficult to estimate source data distribution under different feature spaces. Instead, this project proposes to extract source and target regions from label neighbours, which could (i) induce source and target data distributions, and (ii) facilitate cross- domain alignment. As a result, an adaptable and discriminative target classifier can be developed.

The results of this project will not only contribute additional knowledge to improve the generalisation capability of classifiers and machine learning models, but will also be applicable to the visual recognition software industry. This new approach will support user/developer friendly software deployment. In future, when applying a recognition engine into a new environment that is different from the training environment, the recognition engine could automatically be refined using target domain unlabelled data to adapt to the new environment.