DEEP NEURAL NETWORKS FOR OBJECT-DETECTION AND INSTANCE SEGMENTATION OF MECHANICAL, ELECTRICAL AND PLUMBING COMPONENTS: TRANSFER LEARNING ON RADIATORS AS AN EXAMPLE

Cost estimation and as-built documentation on construction sites require extensive manual work. Combining LiDAR with computer vision offers an effective solution for 5D BIM applications. Therefore, two deep neural networks are trained on a database of over 1000 images of radiators, as example of MEP components. Transfer learning has been applied on SOLOv2 for instance segmentation and on YOLOx for object detection. Given the variety of the image database, the generated models achieve satisfying performance, suitable for the intended applications. The developed models will be used in a scan-to-BIM workflow, enabling the semi-automatic capture of as-built information.