Transfer learning for deep learning-based point cloud tasks in construction scenes

Deep learning (DL)-based point cloud processing has been widely applied to 3D reconstruction in the construction industry. However, DL methods typically require large, fully annotated data sets for effective learning, which can be time-consuming and labor-intensive to produce. This paper proposes a transfer learning method to improve the effectiveness of point cloud tasks for construction scenes. This paper investigates the performance of pretrained models across various DL tasks, algorithms, and transfer learning configurations, specifically fine-tuning and partial transfer learning, and evaluates them in three representative construction scenarios: underground garage data set (UGD); construction site data set (SITE); and pipe system network data set (PSNet). The results demonstrate the following: (1) pretraining on data sets from diverse construction environments improves the mean intersection over union by 8.7%, 8.3%, and 45% on UGD, SITE, and PSNet, respectively, with backbone-only pretraining achieving the best performance; (2) a moderate pretrained sample size combined with a larger training sample size achieves better pretraining results; and (3) pretraining can also improve the accuracy of instance segmentation and point completion.