Efficient Semantic Segmentation via Advanced Prompt Tuning Techniques

Semantic segmentation is a fundamental task in computer vision with critical applications in autonomous driving, medical imaging, and remote sensing. However, optimizing segmentation models for real-world deployment remains challenging due to the high annotation costs and computational demands associated with large-scale datasets. In this paper, we propose a novel approach that synergistically integrates superpixels, active learning, and visual prompt tuning to enhance semantic segmentation efficiency and generalization across diverse domains. Our method leverages superpixels to guide region-based active learning, strategically selecting the most informative areas for annotation, thereby minimizing labeling costs while preserving segmentation quality. Additionally, visual prompt tuning optimizes feature adaptation, reducing both training and inference time. We validate our approach on standard domain adaptation benchmarks, including SYNTHIA, GTAV, and Cityscapes, using SAM-ViT as the backbone. Experimental results demonstrate superior performance over state-of-the-art methods, achieving faster convergence and lower computational overhead while preserving segmentation performance.