Domain Adaptation Through Weak and Self-Supervision for Small Object Segmentation in Construction Site Monitoring

Abstract

Automating construction site monitoring through deep learning-based segmentation presents challenges due to the high cost of pixel-wise annotations. This study proposes a novel weak and self-supervised learning framework to enhance segmentation accuracy while reducing annotation burdens. Utilizing the Segment Anything Model (SAM), we generate high-quality polygon mask labels from bounding boxes, which are refined via self-training. Compared to fully supervised learning models, the proposed method integrates Transfer Learning, Pseudo-Label Refinement, and the Noisy Student technique, improving Mask mean Average Precision (Mask mAP) by 3–56% across four target domains. It is expected to achieve a segmentation Mask mAP of 76.84% while reducing manual annotation time by at least fivefold. Additionally, the proposed method outperformed existing weakly supervised techniques, such as BoxSnake and BoxTeacher, by 26.5% and 29.7%, respectively. These findings contribute to scalable and efficient construction safety monitoring, improving worker safety compliance in real-world applications.

Keywords: Weak Supervision, Self-Supervised Learning, Domain Adaptation, Small Object Segmentation, Construction Safety Monitoring, Pseudo Label Generation, Deep Learning in Construction