| description abstract | The use of automatic building information modeling (BIM) based on point cloud scans is in increasing demand in many engineering applications, such as construction progress monitoring, building renovation, project management, energy simulation, and defect detection. Segmentation-based three-dimensional (3D) modeling approaches using deep learning have been extensively investigated and achieved great performance in recent years. However, segmentation-based methods represent an object as a cluster of points and require further handcrafted steps to convert those points into 3D models. This paper aims to achieve a fully automatic, high-precision method of scan-to-BIM by exploring new 3D object detection networks. CorDet, a corner-aware detector, is proposed for the reconstruction of 3D objects in BIM. Each building object is represented as a class-specific, oriented, and symmetric 3D bounding box. The local features around the corners of an object are incorporated in order to decompose the object location precisely. CorDet can simultaneously learn both object-level and corner-level features through corner-based supervision using deformable convolutions. In experiments on the S3DIS data set, CorDet outperforms state-of-the-art benchmarks with a detection accuracy of 80.5% and a mean intersection of union (mIoU) of 88.9%. The average time spent in modeling a single room is 0.53 s, and this scheme therefore has great potential for many real-time engineering applications. | |
