Hybrid self-supervised learning-based architecture for construction progress monitoring

Automating construction progress monitoring is essential for the timely completion of projects. Computer vision -based construction progress monitoring (CV-CPM) stands out as a promising technology, leveraging 3D point clouds as inputs. Both heuristics-based and learning-based approaches have been explored for identifying building elements. Nevertheless, prevailing supervised methods require project-specific manual labeling, rendering them non-generalizable. This paper introduces a hybrid self-supervised learning architecture named ConPro-NET, which integrates heuristics with learning-based techniques for element identification from con-struction point clouds. The proposed approach conducts unsupervised segmentation through a region-growing -based method, followed by feature extraction using contrastive learning. Contrastive learning matches object pairs to learn their features, which are refined and augmented with handcrafted features based on local geo-metric and visual properties to form the hybrid feature vector. The model demonstrates an overall classification accuracy of 80.86% on the S3DIS dataset and 80.95% on a case study dataset, encompassing the classification of six object classes.