An adaptive density weighting algorithm for semantic segmentation in large-scale oblique photogrammetric point clouds

Large-scale oblique photogrammetry point clouds exhibit significant density variations and scale differences, presenting challenges for urban scene semantic segmentation. Existing methods struggle to balance efficiency and accuracy on non-uniform point clouds. To address this issue, we propose RandLA-ADW, an end-to-end density-adaptive model extended from RandLA-Net. It integrates local density estimation into feature learning via an adaptive density weighting (ADW) module, which dynamically adjusts feature aggregation weights for sparse regions using KNN-based density estimation. A density-aware loss is further introduced to regularize density encoding and enhance robustness. Evaluated on the SensatUrban dataset, RandLA-ADW improves mAcc to 80.3% and mIoU to 67.55%, outperforming RandLA-Net by 8.2 and 7.83 percentage points respectively. It achieves superior segmentation on fine-grained and low-density classes, providing an effective solution for oblique photogrammetry point cloud semantic segmentation.