Automated handling of reflection for elimination of incorrect points and object reconstruction from laser point cloud

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


Mirrors are common in our everyday lives and cause reflection of 3D points captured during laser scanning, necessitating their elimination in data post-processing. However, reflection can also be beneficial for capturing hard-to-reach parts of objects. This study aims to develop an automated solution for handling reflected points encountered during 3D laser scanning of buildings and facilities. Therefore, this research reformulates the mirror detection problem as identifying reflected points, beginning with frame detection. Ten generic deep learning networks/models and existing custom models used for mirror detection in RGB images were investigated for frame detection. The proposed methodology involves identifying frames and categorizing them as picture frames or mirrors using DBSCAN and eliminating or correcting the reflected points when mirrors are detected. The developed method was validated on a dataset containing 50 scans with mirrors and pictures, showed that generic models outperformed custom ones for frame detection, and the proposed method achieved satisfactory results, even with multiple frames in the point cloud.