Geometry-Enhanced Point–Voxel Fusion with Active Learning for Label-Efficient Point Cloud Semantic Segmentation

Abstract

Point cloud semantic segmentation is fundamental for 3D scene understanding and has been widely used in autonomous driving and infrastructure inspection applications. However, its performance is often limited by insufficient representation of local geometric structures and the high cost of point-wise annotation. To address these issues, this paper proposes GeoFuse-AL, a label-efficient segmentation framework that integrates a geometry-guided point–voxel network with multi-cue active learning. The base model, GeoFuseNet, builds on a hybrid point–voxel backbone and incorporates a Local Geometry Prototype Attention module to enhance object boundaries, fine-grained structures, and local geometric patterns. An Adaptive Channel Fusion module is further designed to improve feature interaction between point-level details and voxel-level context. To reduce annotation dependence, a Multi-Cue Diversity Active Sampling strategy combines prediction uncertainty, color-gradient variation, geometric curvature, and feature-space clustering to select informative and diverse samples. Experiments on S3DIS and SemanticKITTI demonstrate that the proposed model achieves mIoU scores of 63.3% and 61.7%, respectively, outperforming several representative methods. Under limited annotation settings, the proposed strategy reaches 99.2% of fully supervised performance with only 15% labeled data on S3DIS and 97.6% with only 5% labeled data on SemanticKITTI. These results demonstrate that GeoFuse-AL improves segmentation accuracy while substantially reducing annotation requirements.