GeoContrast: Geometric knowledge-based contrast learning for industrial point cloud segmentation

Abstract

Process simulation of industrial robots is essential for improving operational accuracy and efficiency. Due to their three-dimensional structure, point clouds show great promise for building high-fidelity simulation environments. However, the absence of semantic information in unstructured point clouds exacerbates the consistency issue between physical entities and the corresponding virtual models. Several factors hinder improvements in the accuracy of point cloud semantic segmentation. These include the variability of process equipment with a high degree of freedom, the inherent complexity of industrial environments, and the scarcity of high-quality annotated data. To bridge this gap, this paper proposes a novel contrast pre-training framework, GeoContrast, tailored for point cloud segmentation in industrial environments. Firstly, a geometric knowledge module generates geometrically consistent features through clustering and multidimensional feature encoding, improving the segmentation accuracy and generalization ability of high degree of freedom entities. Secondly, an Upsampling Attention Module is implemented to efficiently fuse multi-layer features, significantly enhancing the robustness of entity boundary segmentation. Finally, an uncertainty-weighted contrast loss function is proposed that incorporates information entropy to dynamically adjust sample weights, improving the model’s ability to discern entity boundaries in complex industrial environments. The effectiveness of the proposed method was validated using the Actual Welding Scene Scanning (AWSS) dataset and the Stanford 3D Indoor Scene (S3DIS) dataset. Results demonstrated that the proposed method outperformed existing methods, achieving a maximum improvement of 45.2% in a mean intersection over Union (mIoU) on the AWSS dataset using only 1% of the labels. In addition, it also achieved excellent performance on S3DIS, with mIoU of 75.3%.