TS40K: a 3D Point Cloud Dataset of Rural Terrain and Electrical Transmission System

TL;DR

TS40K introduces the first public 3D point cloud benchmark focused on rural European electrical transmission systems, captured by UAV-LiDAR with over $40{,}000$ km of infrastructure and labeled into $22$ classes. The paper provides comprehensive benchmarks for 3D semantic segmentation and 3D object detection, comparing leading models (e.g., Point Transformer V2, KPConv, PV-RCNN) and highlighting the impact of aggressive data imbalance and noisy, inspection-driven annotations. Key findings show that while state-of-the-art methods achieve strong overall performance (e.g., $mIoU$ around $62.3\%$ for segmentation and $mAP$ around $61.23\%$ for detection), they struggle on critical rural elements like power-line towers, underscoring the need for robust learning under imbalance and labeling noise. The work also discusses challenges and future directions, including generalization across diverse rural environments and multimodal fusion, to advance reliable rural infrastructure inspection applications.

Abstract

Research on supervised learning algorithms in 3D scene understanding has risen in prominence and witness great increases in performance across several datasets. The leading force of this research is the problem of autonomous driving followed by indoor scene segmentation. However, openly available 3D data on these tasks mainly focuses on urban scenarios. In this paper, we propose TS40K, a 3D point cloud dataset that encompasses more than 40,000 Km on electrical transmission systems situated in European rural terrain. This is not only a novel problem for the research community that can aid in the high-risk mission of power-grid inspection, but it also offers 3D point clouds with distinct characteristics from those in self-driving and indoor 3D data, such as high point-density and no occlusion. In our dataset, each 3D point is labeled with 1 out of 22 annotated classes. We evaluate the performance of state-of-the-art methods on our dataset concerning 3D semantic segmentation and 3D object detection. Finally, we provide a comprehensive analysis of the results along with key challenges such as using labels that were not originally intended for learning tasks.

Figures (7)

• Figure 1: The TS40K dataset is derived from raw 3D scans illustrated in Figure \ref{['fig:teaser-raw-sample']} and processed into three different sample types: (1) Tower-radius focuses on the towers that support power lines and its environment (Fig. \ref{['fig:teaser-tower-radius']}). Conversely, (2) Power-line samples have power lines as their main focus in the 3D scenes (Fig. \ref{['fig:teaser-power-line']}). Lastly, (3) No-tower samples represent rural terrain where the transmission system is located, excluding supporting towers but potentially including power lines (Fig. \ref{['fig:teaser-no-ts']}). In Figures \ref{['fig:density1']} and \ref{['fig:density2']}, we showcase the semantic class densities of the TS40K dataset. Figure \ref{['fig:density1']} illustrates the class density for each of the aforementioned sample types. In turn, Figure \ref{['fig:density2']} shows the overall class density in the TS40K train and test sets.
• Figure 2: Examples of TS40K raw 3D point clouds. Different semantic classes are labeled with different colors.
• Figure 3: The raw TS40K land strips are partitioned into three sample types: Tower-radius: Encompasses areas around power-line supporting towers (Fig. \ref{['fig:tower-radius']}); Power-line: Focuses on power lines between towers (Fig. \ref{['fig:power-line']}); and No-tower: Represents rural areas without towers but potentially including power lines(Fig. \ref{['fig:no-ts']}). This categorization ensures safety and addresses data imbalance.
• Figure 4: In real-world 3D datasets not specifically tailored for machine learning tasks, noisy labeling can be a significant challenge. In our scenario, within the TS40K dataset, instances of mislabeled 3D elements are apparent. For instance, patches of ground might be incorrectly labeled as supporting towers, and occasional noise and power lines might be mistakenly classified as medium vegetation. This mislabeling can occur due to safety considerations around tower areas and the presence of power lines not connected to the main grid, which may not be properly identified.
• Figure 5: Qualitative results showcasing the performance of Point Transformer V2 (PTV2) wu2022point on the TS40K dataset. In the first row, PTV2 successfully predicts the primary tower in the scene and accurately identifies smaller voltage towers, often overlooked in the ground truth annotations. However, the second row reveals an instance where PTV2 introduces a patch of ground surrounding two towers that was absent in the original labels. This highlights the impact of noisy labels in 3D benchmarks like PTV2.