Data-centric enhancement of site-specific automated construction equipment detection in wide-angle site images

Abstract

Construction equipment detection in wide-angle site images is limited by data scarcity, site-specific variability, and adaptability of existing models. This paper presents a site-specific automated framework from a data-centric perspective to enhance detection performance. Site-specific datasets are generated using zero-shot instance segmentation and depth estimation to create synthetic images that reflect actual site conditions. Object detection models are trained on these datasets, and a slicing-based inference pipeline is integrated to further improve detection performance. Four model configurations are compared: combining two equipment image types (bounding-box and segmented objects) and two synthetic methods (scale-agnostic and scale-aware). The framework improves detection performance by up to 13.72 % over the control group. Requiring minimal human intervention, it offers a reproducible and scalable approach for developing site-specific object detection models, supporting downstream applications such as productivity analysis and safety monitoring.