Development of a Pipe Recognition Network Based on Attention Mechanism

Abstract

Automatic pipe recognition is required in the construction of piping BIM models. Commonly, 3D point clouds are used for 3D piping recognition. However, an enormous amount of calculation was required. Therefore, in this study, we proposed a method for pipe recognition using image segmentation networks based on the deep learning technique. To improve the performance of the recognition network, the Attention Mechanism based Decoder and the Rectangular Atrous Spatial Pyramid Pooling are proposed. Through experiments, the performance of the proposed network was evaluated.