Real-time vision-based defect detection for large-scale on-site earthen additive manufacturing: Annotated dataset and dual-model framework

Additive manufacturing is transforming construction through automated, digitalized workflows, yet the layer-wise process remains prone to defects and geometric deviations during long-duration prints. Vision-based quality monitoring, when supported by domain-specific annotated datasets, enables deep learning models to detect such defects in real time. This paper introduces a publicly available annotated, material-specific dataset focused on large-scale earthen 3D-printing (3DP-E) under outdoor on-site conditions, comprising 2000 images labeled into three crack-related classes. A catalog of 14 visually distinguishable defect types is compiled, prioritizing those most relevant for computer vision-based monitoring. To demonstrate the dataset’s applicability, a dual-model vision framework was implemented, stacking a nozzle-tracking model and a crack-segmentation model within a dynamic region of interest (ROI) that adapts to changing camera positions and printer motion. The framework confirms the feasibility of real-time defect detection and outlines a pathway for integrating vision-based supervision into feedback-driven process control workflows for on-site additive manufacturing. • Introduce 3DPE-Crack22, a dataset for defect detection in additive manufacturing. • 2000 annotated images: three crack categories under varying lighting conditions. • A catalog of 14 distinguishable defect types for vision-based monitoring. • Dual-model framework for nozzle tracking and crack segmentation. • Demonstrate vision-based monitoring for quality control in additive manufacturing.