Understanding Point Clouds of Electrical Substations Using a Multi-View Multi-Scale Fusion With Transformer

We propose a deep learning-based network model, relying on multi-view multi-scale fusion with transformer, for understanding semantical information of electrical substation point clouds. Our method is inspired by human recognition of 3D objects using information from multiple perspectives and with multiple levels of details, and it is constructed by a dynamic integration of multi-view 3D morphological representation and multi-scale 3D geometric characterization. Additionally, we leverage the channel and spatial attention mechanism to capture the relationship between morphological representations at multiple views and utilize the self and cross attention mechanism to understand the relationship between geometric characterization at multiple scales. To verify our model, experiments were conducted based on our substation point clouds dataset and two benchmark datasets. Our model shows a better performance than the state-of-the-art methods, and it achieves an overall accuracy of 91.15% on our dataset, 93.30% on ModelNet40, and 92.80% on ModelNet40-C, indicating its effectiveness and robustness.