Active learning enhances annotation efficiency by focusing on the most informative samples, reducing reliance on extensive human input. Existing semantic segmentation methods often overlook the broader structural patterns in images and the potential of advanced pre-trained models. To address these gaps, we propose Entity-Superpixel Annotation (ESA), a novel active learning strategy leveraging a class-agnostic mask proposal network and superpixel grouping to capture local structures. ESA prioritizes high entropy superpixels, selecting key entities in target domain images to balance efficiency and representation. This annotator-friendly approach achieves substantial gains, requiring only 40 clicks per annotation compared to 5000 in traditional methods, reducing click costs by 98%, and improving performance by 1.71%. By integrating structural cues and minimizing annotation effort, ESA outperforms existing pixel-based methods, offering a more effective and efficient solution for semantic segmentation.