Research map / Point clouds / Segmentation and detection
Generated and Pseudo Content guided Prototype Refinement for Few-shot Point Cloud Segmentation
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.
Proposes GPCPR, a few-shot 3D point cloud segmentation framework that refines support prototypes with LLM-generated class/differential descriptions and class-specific pseudo-query context, plus a dual-distillation loss.
Problem
Few-shot 3D point cloud semantic segmentation suffers from low-quality prototypes due to limited semantics in the support set and class information bias between support and query sets.
Approach
Uses GPT-3.5-generated diverse and differentiated class descriptions (encoded by CLIP text encoder with adapters) to enrich prototypes via text-to-prototype cross-attention compressors, then generates pseudo query masks to extract class-specific query context for query-specific prototypes, with a self-distillation loss between early and deep prototypes and pseudo/final logits.
Key results
- Surpasses SOTA DPA by 1.19%-12.10% mIoU on S3DIS across settings.
- Surpasses SOTA DPA by 3.96%-13.75% mIoU on ScanNet across settings.
- Ablation: GCPR (both description types) improves QGPA baseline by 10.62% on S3DIS 2-way 1-shot; full model reaches 75.74% mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GPCPR instead of DPA | Few-shot point cloud segmentation · S3DIS 2-way 1-shot mean | mIoU (%): 70.19 -> 75.74 | Helped |
| use GPCPR instead of DPA | Few-shot point cloud segmentation · ScanNet 2-way 1-shot mean | mIoU (%): 62.90 -> 73.93 | Helped |
| add diverse descriptions D only (ablation) | QGPA baseline · S3DIS 2-way 1-shot mean | mIoU (%): 61.02 -> 69.70 | Helped |
| add differentiated descriptions D′ only (ablation) | QGPA baseline · S3DIS 2-way 1-shot mean | mIoU (%): 61.02 -> 67.55 | Helped |
| add PCPR only (ablation) | QGPA baseline · S3DIS 2-way 1-shot mean | mIoU (%): 61.02 -> 68.45 | Helped |
| add both diverse and differentiated descriptions (ablation) | QGPA baseline · S3DIS 2-way 1-shot mean | mIoU (%): 61.02 -> 71.64 | Helped |
| add prototype distillation LT P (ablation) | GPCPR without LT P · S3DIS 2-way 1-shot mean | mIoU (%): 71.64 -> 72.21 | Helped |
| add prototype distillation LQP (ablation) | GPCPR without LQP · S3DIS 2-way 1-shot mean | mIoU (%): 72.21 -> 72.61 | Helped |
| add pseudo-prediction distillation LQM (ablation) | GPCPR without LQM · S3DIS 2-way 1-shot mean | mIoU (%): 72.21 -> 73.59 | Helped |
| replace gpt-3.5-turbo with gpt-4o-mini (ablation) | GPCPR · S3DIS 2-way 1-shot mean | mIoU (%): 75.74 -> 73.88 | Hurt |
Surprising findings
- gpt-4o-mini yielded slightly lower results than gpt-3.5-turbo (73.88 vs 75.74 mIoU), attributed to longer, noisier paragraphs.
- Differentiated class-pair descriptions helped more than diverse single-class descriptions (7.43% vs 6.53% gain over baseline).
Limitations
- Adds LLM and CLIP overhead; offline generation takes 30-67 minutes depending on dataset.
- Differentiated descriptions scale O(N^2) with number of classes.
- Performance depends on quality of LLM descriptions and pseudo masks; evaluated only within-scene unseen classes.
Experimental setup
Datasets: S3DIS ScanNet
Benchmarks: S3DIS ScanNet
Metrics: mIoU
Baselines: ProtoNet AttMPTI BFG SCAT QGPNet 2CBR QGE QGPA DPA
Builds on: QGPA DGCNN CLIP masked average pooling
Closest papers by meaning
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.88)
- DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation (ICML 2025 · similarity 0.88)
- Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model (CVPR 2025 · similarity 0.88)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.88)
- Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.