Comparative Study of Structure from Motion on Construction Site

For progress monitoring in the construction industry, understanding the state of construction sites is crucial. However, traditional manual inspection methods are labor-intensive and time-consuming. To address these challenges, various methods for creating 3D models of job sites have been explored. Professional equipment such as LiDAR and laser scanners offer the most accurate means of generating point clouds (PCDs) and constructing 3D models. However, these tools are expensive, cumbersome, and often impractical for frequent use in dynamic construction environments. Recently, with the advancement of deep learning, 3D reconstruction techniques have been extensively studied and applied across various fields. Among these, Structure from Motion (SfM) stands out as a method capable of generating PCDs and estimating camera poses. Based on advancing capabilities of SfM, many research has been conducted to measure progress monitoring in construction field. However, most studies that have utilized SfM for progress monitoring have acquired a large number of images and ensured significant overlap in input data to enhance the robustness of the 3D model. While this approach provides a highly accurate 3D reconstruction, the image acquisition process itself introduces additional labor-intensive tasks. Therefore, this study aims to adhere to the fundamental nature of 3D reconstruction by evaluating the performance of various SfM models using only 26 images captured from a brief video recording at a construction site. The findings aim to evaluate the applicability of various SfM technologies with limited data, in real-world construction scenarios and finally provide insights into their potential and future directions.