Accurate concrete spalling segmentation from bounding box supervision using Segment Anything

Abstract

Accurate pixel-level segmentation of concrete spalling has been severely hampered by the prohibitive cost of manual annotation. This paper investigates how accurate pixel-level defect segmentation can be achieved using only low-cost weakly supervised bounding box annotations. A three-stage framework is proposed to generate and refine pseudo-masks from bounding boxes using the Segment Anything Model (SAM), dynamic self-correction, and inference-time fusion. The proposed method outperformed existing techniques by over 10% in F1 score on a large-scale spalling dataset. These findings establish the economic viability of deploying scalable automated inspection systems by drastically reducing data annotation costs, providing a practical and scalable pathway for spalling assessment.