Partial points segmenter: A unified framework for partially category-supervised point cloud segmentation

In many real-world 3D segmentation scenarios like infrastructure inspection and asset management, only a subset of semantic categories are of interest, rendering full-scene annotation both unnecessary and expensive. However, existing methods are primarily designed for fully supervised settings with dense labels across all categories. While weakly supervision alleviates annotation burden, it still assumes coverage over the entire class set, limiting its applicability to only interest categories are labeled. To fill this gap, we propose Partial Points Segmenter (PPS), a unified framework for 3D point cloud semantic segmentation when under partial category supervision. PPS introduces a prototype-guided classification module and a multi-objective training strategy comprising prototype contrastive loss, orthogonality regularization, and adaptive focal foreground loss. These components collaboratively enhance category discrimination under sparse supervision. Extensive experiments across three benchmarks show that PPS significantly outperforms baselines, achieving up to a 19% improvement in mIoU, demonstrating strong generalization in task-specific, partially labeled 3D scenes. The source code and data are available at https://github.com/Lionelsy/PPS.