3D deep learning enhanced void-growing approach in creating geometric digital twins of buildings

The challenge this paper addresses is how to automatically generate geometric digital twins of the indoor environment of buildings. Unlike most previous research that starts with detecting planes in the point cloud and only considers the geometric information, the proposed ”void-growing” approach is a full-automatic approach that starts with detecting void space inside rooms, considering geometric information, as well as semantic information predicted from deep learning. Then based on the detected room spaces, structural elements, as well as doors and windows, are extracted. The method can work in (1) rooms with complex structures like U-shape and L-shape, (2) rooms with different ceiling heights, and (3) rooms under a high occlusion level. Compared with previous studies that mainly only use geometric information, the approach also focuses on how to select useful information predicted by deep learning. This study used existing state-of-the-art deep learning architecture for the segmentation task in the proposed approach. By taking useful semantic information into consideration, the proposed approach performs better in creating geometric digital twins of buildings.