This paper presents an active learning framework for robust object detection in dynamic construction environments, addressing the challenges of limited labeled data and high annotation costs. The framework integrates YOLOv10 with a weighted, adaptive uncertainty–diversity sampling strategy and employs transfer learning to mitigate cold-start issues and accelerate model convergence. The adaptive fusion mechanism dynamically weights multiple uncertainty measures (classification confidence, class entropy, and bounding box variance) while incorporating sample diversity to prioritize the most informative data. Experiments achieved mAP50 of 0.885 and mAP50–95 of 0.730 by Cycle 8. Rare classes showed notable gains: Circular ducts improved from 0.25 to 0.60 mAP50, and Drywall panels from 0.35 to 0.65. The approach reduced labeling effort by 30–40% compared to random sampling, showing its potential to improve the development of construction element recognition algorithms, including small items that are often not visible, such as safety equipment, building parts, and construction tools. • Weighted adaptive uncertainty-diversity sampling reduces construction site labeling effort by 30–40% versus random selection. • Dynamic multi-component uncertainty aggregation balances classification confidence, entropy, and localization precision. • Category-wise difficulty calibration improves rare class detection: Circular ducts +140%, Drywall panels +86% mAP50 gain. • Model-agnostic framework achieves mAP50 = 0.885, mAP50–95 = 0.730 on imbalanced real-world construction dataset (17 classes). • Transfer learning from COCO resolves cold-start issues, enabling effective training with limited initial labeled data.
Mannem et al. (Tue,) studied this question.