Abstract
Active Domain Adaptation (ADA) queries the label of selected target samples to help adapting a model from a related source domain to a target domain. It has attracted increasing attention recently due to its promising performance with minimal labeling cost. Nevertheless, existing ADA methods have not fully exploited the local context of queried data, which is important to ADA, especially when the domain gap is large. In this paper, we propose a novel framework of Local context-aware Active Domain Adaptation (LADA), which is composed of two key modules. The Local context-aware Active Selection (LAS) module selects target samples whose class probability predictions are inconsistent with their neighbors. The Local context-aware Model Adaptation (LMA) module refines a model with both queried samples and their expanded neighbors, regularized by a context-preserving loss. Extensive experiments show that LAS selects more informative samples than existing active selection strategies. Furthermore, equipped with LMA, the full LADA method outperforms state-of-the-art ADA solutions on various benchmarks. Code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2208.12856
