Context-Guided Active Domain Adaptation for Blended Target Domain

Abstract

Active domain adaptation (ADA) queries the labels of a limited number of selected target samples to help transfer knowledge from a source domain to a target domain. However, most ADA methods are still mainly oriented to a single target domain and not applicable for the blended target domain. To tackle these issues, we propose a concise and effective approach named Context-Guided Active Domain Adaptation (CGDA) to achieve active blended-target domain adaptation (BTDA). CGDA captures spatial context and local context to effectively use and understand context information to significantly improve the performance of active BTDA. First, we capture the spatial context relations of the target data through a masked image aware (MIA) module and then adapt the model to the inputs of a blended target domain through a blended feature augment (BFA) module. Furthermore, we utilize the local inconsistency of model predictions and uncertainty to design a selection criterion for selecting samples with more abundant local context information. Experiments on several BTDA datasets show that the performance of CGDA is significantly better than existing BTDA methods.