Abstract
Accurate documentation of underground infrastructure is critical for preventing accidental damage during excavation, which costs billions annually. OpenTrench3D is the first publicly available 3D point cloud dataset specifically designed for the semantic segmentation of underground utilities in open trenches. By utilizing smartphone photogrammetry and state-of-the-art deep learning, this work demonstrates a scalable approach to automating utility mapping and documentation.
Introduction
Accidental strikes on underground utilities—such as water, gas, and fiber optic lines—result in significant global economic losses, estimated at $30 billion in the US in 2019 alone. These incidents are primarily driven by inaccurate or outdated utility maps. While surveying utilities in open trenches is the gold standard for accuracy, the manual nature of current documentation processes often leads to incomplete records, as surveyors frequently ignore secondary exposed utilities to save time.
Existing 3D datasets, like ScanNet or SemanticKITTI, focus on indoor or urban surface environments, leaving a significant gap in data for the underground domain. OpenTrench3D addresses this by providing high-quality, annotated 3D data of exposed utilities, enabling the training of AI models to recognize infrastructure in complex, real-world trench environments.
Methodology
The dataset was generated using a cost-effective close-range photogrammetry approach. Workers captured video of open trenches using standard smartphones, which were then processed via Structure from Motion (SfM) and Multi-View Stereovision (MVS) to create dense 3D point clouds.
To evaluate the dataset, three state-of-the-art point cloud segmentation models were benchmarked: PointNeXt, PointVector, and PointMetaBase. The study particularly emphasized transfer learning, examining how models pre-trained on water utility data could adapt to district heating utilities with minimal additional samples.
Dataset Description
OpenTrench3D consists of 310 fully annotated point clouds, totaling approximately 528 million points. The data was collected across seven distinct geographic areas in Denmark, covering five water utility areas and two district heating areas. Points are classified into five categories:
- Trench: Background soil and trench walls.
- Main Utility: The primary pipe being installed or serviced.
- Other Utility: Existing pipes belonging to other infrastructure owners.
- Inactive Utility: Abandoned or out-of-service pipes.
- Misc: Tools, debris, and other non-utility objects.
Results and Discussion
The benchmarking results highlighted several key findings regarding the performance of deep learning models on underground utility data:
- Transfer Learning Efficiency: Fine-tuning a pre-trained model on just a single sample from a new geographic area significantly improved performance compared to a general pre-trained model. Models trained from scratch on small subsets performed significantly worse, proving the value of the OpenTrench3D dataset for pre-training.
- Model Performance: PointVector generally outperformed other architectures, likely due to its vector-oriented abstraction being well-suited for the linear and cylindrical geometries of pipes. In water utility areas, PointMetaBase achieved a mean Intersection-over-Union (mIoU) of 90.5% when excluding the challenging "Inactive" class.
- Cross-Domain Robustness: Models successfully transferred "pipeness" features between different utility types (e.g., from blue water pipes to black insulated heating pipes), achieving a 76.6% mIoU on heating areas with only 10 fine-tuning samples.
- The Inactive Utility Challenge: Distinguishing between active and inactive pipes of the same type remains the most difficult task due to their near-identical visual appearance, often requiring historical map data or expert intuition to resolve.
Conclusion and Future Work
OpenTrench3D provides a foundational resource for automating underground utility mapping. While the models show strong performance, challenges remain regarding class imbalance (trench points dominate) and the visual ambiguity of inactive pipes. Interestingly, a "negative trend" was observed where performance sometimes dipped when increasing the fine-tuning set from 10 to 29 samples, likely due to hyperparameter optimization challenges in small-scale fine-tuning. Future research may explore multi-modal approaches, integrating 3D point clouds with historical GIS records to improve classification accuracy for visually similar utilities.
Summary updated