Reproduce “Rethinking Few-shot 3D Point Cloud Semantic Segmentation” on OpenResearch

About this paper

Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, Serge Belongie

This research reveals critical flaws in existing benchmarks for few-shot 3D point cloud semantic segmentation (FS-PCS) and introduces COSeg, a novel model based on "correlation optimization." It establishes a new state-of-the-art, achieving mIoU improvements of up to 6.82% on S3DIS and 6.58% on ScanNet datasets under a rigorously standardized evaluation setting.

Key insights

• Existing FS-PCS benchmarks were fundamentally flawed due to foreground leakage and overly sparse point distributions, leading to inflated performance metrics that did not reflect true few-shot generalization.
• Directly optimizing the correlations between query points and class prototypes, rather than merely refining features, provides a more effective paradigm for few-shot learning in 3D point clouds.
• Dynamically maintained non-parametric prototypes for base classes can effectively calibrate background correlations, reducing the model's bias towards seen categories and improving novel class segmentation accuracy.

Summary by alphaXiv.