Research map / Point clouds / Segmentation and detection

Rethinking Few-shot 3D Point Cloud Semantic Segmentation

CVPR 2024 · Point clouds › Segmentation and detection · approach: Point cloud semantic segmentation · 48 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Identifies foreground leakage and sparse sampling flaws in the few-shot 3D point cloud segmentation (FS-PCS) benchmark, introduces a corrected 20,480-point setting, and proposes COSeg, which refines class-specific multi-prototypical correlations with hyper correlation augmentation and base prototype calibration.

Problem

Current FS-PCS benchmarks leak foreground information through non-uniform point sampling and use only 2,048 points per scene, which inflates results and limits semantic information, so they do not reflect true model performance or real-world conditions.

Approach

Shift from feature optimization (refining support/query features for prototypes) to correlation optimization: refine per-query-point correlations to all class prototypes with linear-attention-based HCA, and calibrate background correlations using non-parametric momentum-updated base prototypes (BPC) to counter base-class susceptibility.

Key results

• On S3DIS 1-way 1-shot in the corrected setting, COSeg reaches 47.21 mIoU, a 6.82-point gain over the second-best QGE.
• Removing foreground leakage drops previous methods' mIoU substantially, e.g., the best S3DIS 5-shot result falls from 81.80% to 45.52% for QGPA.
• Ablation on S3DIS: moving from feature to correlation optimization raises 1-shot mIoU from 30.67 to 39.93; adding HCA and BPC reaches 47.21.

Measured effects reported in the paper

Surprising findings

• Previous FS-PCS methods' high reported scores largely rely on density differences created by foreground leakage; mIoU drops of about 26-28 points on average after correction.
• Non-uniform sampling in the existing FS-PCS protocol leaks foreground class identity via point density.

Limitations

• Results are reported on S3DIS and ScanNet only, with 1-way and 2-way settings.
• The paper's correction relies on a specific sampling protocol; generality to other sampling schemes is not tested.

Experimental setup

Datasets: S3DIS ScanNet

Benchmarks: S3DIS ScanNet

Metrics: mIoU

Baselines: AttMPTI QGE QGPA

Builds on: attMPTI Stratified Transformer linear attention prototypical networks

Closest papers by meaning

• Multimodality Helps Few-shot 3D Point Cloud Semantic Segmentation (ICLR 2025 · similarity 0.92)
• Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model (CVPR 2025 · similarity 0.89)
• Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · similarity 0.89)
• DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation (ICML 2025 · similarity 0.89)
• Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.89)
• Generated and Pseudo Content guided Prototype Refinement for Few-shot Point Cloud Segmentation (NeurIPS 2024 · similarity 0.88)