MACO: An Open-Image Data Set with Multiannotation Strategies for Building Material Counting

Abstract

Building material management is fundamental to engineering projects. Currently, the counting of building materials is primarily conducted manually, a process that is time-consuming and prone to errors. Recent advances in computer vision and deep learning have significantly facilitated automation in the construction industry. However, the effectiveness of deep learning approaches hinges heavily on large-scale data sets with accurate, manually annotated data. Existing construction-domain image data sets are ill-suited for material counting tasks, and this lack of large-scale, publicly available data sets has become a major barrier to progress in this area. To address this gap, this study introduces and publicly releases a new large-scale image data set, called Material Counting in Construction (MACO), collected directly from construction sites. The MACO data set comprises 4,426 images and 563,595 annotated objects, covering seven common building materials and diverse real-world construction scenarios. To enhance the utility of MACO and to support more downstream tasks, three popular annotation techniques were employed: horizontal bounding boxes (HBBs), oriented bounding boxes (OBBs), and segmentation masks. This data set is the first large-scale, multimaterial, multiannotation, and high-density open data set (with 127.3 annotations per image) among existing construction data sets. The validity of MACO is confirmed through benchmarking with two state-of-the-art one-stage object detection algorithms, achieving a maximum mean average precision at an intersection over union (IoU) threshold of 0.5 (mAP50) of 91.6%. This provides a robust benchmark for method selection in similar tasks. In addition, to compare the one-material–one-model and multimaterial–one-model paradigms, experiments were conducted on models trained on single versus multiple materials. Results indicated that the multimaterial counting model exhibits performance degradation due to cross-material feature interference, suggesting that developing a generalized detection and counting model requires further research. Overall, MACO is designed to advance intelligent construction site management, including material detection and counting, specification measurement, and robotic grasping and assembly.