Efficiency is the rule: Domain adaptive semantic segmentation with minimal annotations

Abstract

Active domain adaptation aims to select a few yet informative samples of the target domain for manual annotation to improve model performance. However, a critical observation in our research is the less-than-ideal domain alignment of existing active domain adaptive semantic segmentation (ADASS) methods. Specifically, they only measure the complementarity between target samples and the source domain but neglect the degree of domain shift in the active sample selection process. Furthermore, the impact of hard samples on domain alignment and model discriminative ability is underestimated. To tackle these issues, we propose a framework that contains a novel main-sub anchor modeling method and Confusing Sample Selection (CSS) and Offset Sample Selection (OSS) strategies. While improving the model performance of the ADASS task, we also consider that there remains a substantial resource demand. To solve this issue, we introduce the Instance Assignment Module (IAM). Extensive experiments on GTAV $\to$ Cityscapes and SYNTHIA $\to$ Cityscapes benchmarks, demonstrate that our method sets a new standard in both weakly supervised domain adaptive semantic segmentation (WDASS) and ADASS tasks, achieving the optimal trade-off between annotation cost and model performance.