A Teacher–Student Point Cloud Segmentation Framework for Challenging Features

Existing point cloud semantic segmentation models tend to favor classes with abundant points and salient structures, while sparse, small-scale, and confusable targets remain insufficiently represented. To address this issue, we propose an offline teacher–student point cloud segmentation framework for challenging features. The framework uses challenging-class-centered sampling to strengthen the teacher’s representation of hard-to-learn targets and their local contexts, and generates class-response priors that are point-wise aligned with the original point cloud. A Point Foundation Adapter (PFA) then selectively injects teacher priors at the feature level, while challenging-point selective knowledge distillation imposes targeted constraints at the output level. Experiments on WHU3D and Toronto3D show consistent improvements across different teacher–student backbone combinations. For Sonata–OA-CNNs, the mIoU and Hard mIoU increase from 45.54% and 22.58% to 51.91% and 32.84% on WHU3D, and from 73.03% and 55.41% to 80.73% and 68.51% on Toronto3D, respectively. These results demonstrate that specialized teacher learning and selective prior transfer improve challenging-class segmentation across different backbone architectures, with teacher priors supporting both student training and inference.