Researchers from the University of Applied Sciences Western Switzerland (HES-SO) and the University of Lausanne have developed a new dataset aimed at improving the maintenance and management of power line infrastructure. The team, led by Antoine Carreaud and including Shanci Li, Malo De Lacour, Digre Frinde, Jan Skaloud, and Adrien Gressin, has created GridNet-HD, a multi-modal dataset that combines high-density LiDAR data with high-resolution oblique imagery to enhance the accuracy of 3D semantic segmentation for overhead electrical infrastructures.
GridNet-HD is a comprehensive dataset that includes 7,694 images and 2.5 billion LiDAR points, all annotated into 11 different classes. This dataset is designed to provide a detailed and accurate representation of power line assets, which is crucial for effective maintenance and management. The dataset also includes predefined splits and mean Intersection over Union (mIoU) metrics, which are essential for evaluating the performance of different models.
The researchers have provided baseline models for both unimodal (LiDAR-only and image-only) and multi-modal fusion approaches. The results show that fusion models, which combine both LiDAR and image data, outperform the best unimodal baseline by +5.55 mIoU. This highlights the complementarity of geometry and appearance in accurately identifying and classifying power line assets.
One of the key advantages of GridNet-HD is that it is the first public dataset to jointly provide high-density LiDAR and high-resolution oblique imagery with 3D semantic labels for power-line assets. This makes it a valuable resource for researchers and industry professionals working on improving the maintenance and management of power line infrastructure.
The dataset, along with the baseline models and codes, is available on the Hugging Face platform. This open-access approach ensures that the benefits of this research can be widely shared and applied in practical settings.
In summary, GridNet-HD represents a significant advancement in the field of power line infrastructure management. By providing a high-resolution, multi-modal dataset, the researchers have enabled more accurate and efficient maintenance and management of power line assets, which is crucial for ensuring the reliability and safety of our electrical grids.
The research was published in the IEEE International Conference on Computer Vision (ICCV) 2023.
This article is based on research available at arXiv.