Training-free 3D point-cloud segmentation pipeline recovers 33% of supervised gap

Generalized few-shot 3D point-cloud segmentation (GFS-PCS) traditionally requires extensive training on base classes and few-shot support for novel classes. A new paper by Gah and Owusu introduces a fully training-free alternative that achieves competitive performance using only off-the-shelf vision-language models. The pipeline pairs RegionPLC, a frozen 3D vision-language model providing dense semantic priors, with SAM3, a promptable concept segmenter lifted from posed RGB views. The core innovation is cross-view consistency: a point is labeled as novel only when a majority of RGB views agree on its novel class, eliminating the need for any training, 3D labels, or even few-shot support examples.

On the ScanNet200 benchmark, this approach improves novel-class mIoU by 2.6 points over the dense prior alone while maintaining base-class accuracy within 0.5, recovering roughly one-third of the gap to fully supervised state-of-the-art methods. More impressively, on the harder ScanNet++ dataset—where the dense prior is weak on novel classes—the pipeline nearly doubles novel mIoU from 16.2 to 31.9 with only a 1.7-point base accuracy drop, raising the harmonic mean from 21.5 to 31.1. Interestingly, the authors show that adding few-shot support actually degrades performance, making the method zero-shot in practice. This work opens the door to scalable, annotation-free 3D scene understanding.