A structure‐oriented loss function for automated semantic segmentation of bridge point clouds

Focusing on learning-based semantic segmentation (SS) methods for bridge point cloud data (PCD), this study proposes a structure-oriented concept (SOC) with training focused on the spatial distribution patterns of bridge components, including both the horizontally absolute location of each component and its vertically relative position compared with other components. Then a structure-oriented loss (SOL) function, which embodies the core of SOC, is defined accordingly, and it is compared to five cutting-edge loss functions on a collected bridge PCD dataset. In contrast to the limitations of other loss functions, SOL significantly improves the overall evaluation metrics of overall accuracy (6.53%) and mean intersection over union (mean IoU: 8.67%). The IoU of the category “others” is improved by 8.44%, which is very important for automating the time-consuming denoising process. Furthermore, the demonstrated robustness of SOC and SOL reveal great potential to improve the performance of other SS models.