Optimizing 3D reconstruction: Application of a vision foundation model for dimensional measurement

Abstract

This study applies a vision foundation model to optimize the Point Cloud Data (PCD) reconstruction pipeline in digital photogrammetry, aiming to overcome the challenges of component PCD acquisition in Small and Medium Construction Projects (SMCP). A hardware device is designed and assembled for overlapping image acquisition. In order to alleviate computational and storage burdens, a vision foundation model is utilized for Region of Interest (ROI) selection before PCD reconstruction, for which a set of prompt optimization methods based on image similarity is proposed. The scale of the reconstructed PCD is calibrated based on the camera pose, enabling precise dimension measurement of the PCD. Two case studies confirm the effectiveness of the optimized method for dimension measurement, with tolerance below 2 mm, corresponding to 6.7 times the pixel accuracy. Furthermore, this method demonstrates substantial improvements in image storage, decreasing by over 84 %, and decreases of over 55 % in PCD storage and over 50 % in the computational time required for reconstruction. These outcomes underscore the practicality and efficiency of the introduced framework in addressing the challenges posed by batch component dimension measurement.