In recent years, deep-learning-based visual semantic segmentation tasks have made significant advancements in performance. However, training such deep learning models requires a substantial amount of point-level or pixel-level annotations, which is time-consuming and labour-intensive to acquire. Active learning is an effective learning scheme that enables neural networks achieving maximum performance using a small annotation budget. Recently, many visual semantic segmentation methods have employed active learning strategies to promote promising performance with limited annotations. This paper provides a systematic categorization and comprehensive review of active learning strategies in visual semantic segmentation. First, we briefly introduce the concepts of visual semantic segmentation and active learning, and elicit the significance of active learning in visual semantic segmentation. We then discuss active learning methods for 2D image and 3D point cloud semantic segmentation tasks, respectively. For each task, we provide a detailed analysis of the approaches across three levels of query granularity: image (or point clouds) level, region (or superpoint) level, and pixel (or point) level. Finally, we present the challenges and potential future research directions for active learning in this field.
Xu et al. (Mon,) studied this question.