Session: 02-05-01: Artificial Intelligence and Data-Driven Design 2

Paper Number: 196016

Weakly Supervised Semantic Segmentation for Mechanical and Piping Systems Reconstruction 

The accurate extraction and identification of Mechanical, Electrical, and Plumbing (MEP) components from raw 3D laser scans is a foundational step for reverse engineering and data-driven design. Semantic segmentation serves as the core computer vision methodology for this task, assigning meaningful class labels to unorganized point clouds. However, mechanical systems—such as HVAC ductwork, fluid piping, and structural supports—exhibit intricate, elongated geometries and severe spatial occlusions. Training robust deep learning models for point cloud semantic segmentation currently demands massive amounts of densely annotated data. The manual point-level labeling of such spatially entangled mechanical networks is economically prohibitive, resulting in a severe lack of large-scale benchmark datasets in the engineering domain. This study addresses this critical bottleneck by providing comprehensive real-world datasets and proposing a highly data-efficient weakly supervised semantic segmentation framework to facilitate automated geometric feature extraction under extreme label sparsity.

To establish a robust foundation for evaluating data-driven segmentation algorithms, this paper introduces two complementary, large-scale datasets acquired via high-precision terrestrial laser scanning. The Underground Garage Dataset (UGD) contains over 185 million points, capturing expansive subterranean facilities where complex mechanical utility networks intersect with massive structural elements. To further isolate mechanical features, this study introduces the Building MEP Dataset (BMD), comprising approximately 92.1 million points focused exclusively on multi-system layouts. The BMD explicitly captures the geometric complexity of ventilation ducts, water pipes, cooling units, and electrical conduits across diverse spatial configurations. Together, these datasets exhibit the geometric irregularities, dense proximities, and severe class imbalances typical of real-world mechanical engineering projects, providing a challenging and highly realistic testbed for advancing data-efficient point cloud semantic segmentation.

Recognizing the prohibitive cost of densely annotating complex mechanical layouts, this study develops a weakly supervised semantic segmentation framework engineered to operate under extreme data sparsity. Instead of relying on full supervision, the proposed method utilizes transfer learning and an iterative teacher-student self-training strategy to maximize the extraction of semantic features from vast amounts of unlabeled points. To address the specific spatial characteristics of mechanical components, an adaptive neighborhood querying module is integrated into the architecture. Unlike standard fixed-radius queries, this mechanism dynamically adjusts its receptive field based on local geometric variations, effectively distinguishing the semantic boundaries of tubular pipes from adjacent flat ductwork or planar walls. Additionally, a class-aware pseudo-labeling strategy ensures that the model robustly segments rare mechanical supports without being biased toward dominant volumetric structures during the iterative label propagation process.

Extensive experiments were conducted to evaluate the proposed framework's semantic segmentation capability under simulated sparse annotation conditions. The results indicate that the framework achieves an overall accuracy of 93.50% and a mean Intersection over Union of 81.36% on the BMD dataset using merely 0.01% of the labeled points. On the UGD dataset, the method reaches 97.92% overall accuracy with only 0.02% manual annotation. Compared to conventional fully supervised segmentation architectures, this weakly supervised approach maintains highly competitive fidelity in extracting mechanical semantics while drastically reducing the overall manual annotation workload by approximately 75% to 80%. Ultimately, this paper contributes essential large-scale datasets and a highly efficient algorithmic pipeline to the engineering community, demonstrating that weakly supervised semantic segmentation is a highly viable and scalable solution for the automated reverse engineering of complex mechanical environments.

Presenting Author: Hongzhe Yue Southeast University

Presenting Author Biography: Hongzhe Yue is currently a Ph.D. candidate at Southeast University, China, and a visiting Ph.D. student at the National University of Singapore. His research primarily focuses on the intersection of artificial intelligence and the reverse engineering of complex mechanical environments. Specifically, he specializes in deep learning-based 3D point cloud processing and weakly supervised semantic segmentation for Mechanical, Electrical, and Plumbing (MEP) systems. By developing data-efficient frameworks and geometric extraction algorithms, this research aims to facilitate the automated modeling, spatial analysis, and digital twin generation of large-scale mechanical and piping networks.

Authors:

Hongzhe Yue Southeast University
Qian Wang Southeast University