Advanced Image-Based Methods for 3D Semantic Segmentation of Lidar Point Clouds in Electrical Infrastructure Applications

This study explores AI methods for 3D semantic segmentation relying solely on image data. Using a dataset collected over electrical infrastructures in Switzerland, we evaluate and compare four image-based approaches: majority voting of classifications, logit-based aggregation, distance-weighted logits, and a combination of logits, distance weighting, and depth maps. The results demonstrate that image-only methods can achieve competitive performance, closely approaching the state-of-the-art results achieved with point cloud-based models, while offering the advantages of reduced complexity and cost. Nonetheless, these methods depend on precise alignment between images and point clouds, as well as an effective reprojection process from 2D to 3D.Our experiments also reveal a complementarity between image-based and point cloud-based methods, highlighting the potential for future research on multimodal fusion techniques to fully leverage the strengths of both modalities.