Research map / Detection and segmentation / Open-set and few-shot detection
No Time to Train: Empowering Non-Parametric Networks for Few-shot 3D Scene 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 Seg-NN, a training-free non-parametric point-cloud encoder for few-shot 3D scene segmentation, and Seg-PN, which adds a lightweight QUEST query-support transfer module (0.24M params) trained without pre-training.
Problem
Few-shot 3D scene segmentation pipelines pre-train a learnable encoder on seen classes and then fine-tune episodically, which costs substantial training time and creates a domain gap on unseen classes.
Approach
Replace the learnable encoder with a non-parametric (training-free) Point-NN-style encoder using trigonometric positional encodings and randomly sampled low-frequency hand-crafted filters, plus a similarity-based prototype segmentation head; add a small QUEST module that transfers support prototypes toward the query domain via cross- and self-correlation.
Key results
- Seg-PN achieves 66.41% (2-way 1-shot, S0/S1 avg) on S3DIS, +4.19% average mIoU over prior SOTA, and +7.71% on ScanNet, with 0.24M parameters.
- Seg-NN, with zero learnable parameters, reaches 49.53% on S3DIS 2-way 1-shot, outperforming DGCNN (37.57%) and ProtoNet (49.19%) without training.
- Seg-PN reduces total training time to 0.5 h versus 4.7 to 9.5 h for prior methods (over 90% reduction).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Seg-PN with QUEST | PAP3D · S3DIS, average across four few-shot tasks | mIoU: +4.19% | Helped |
| add Seg-PN with QUEST | PAP3D · ScanNet, average across four few-shot tasks | mIoU: +7.71% | Helped |
| remove pre-training and use the non-parametric Seg-NN encoder | few-shot 3D segmentation · S3DIS | training time: 4.7 h -> 0.5 h | Helped |
| add Seg-NN’s position-and-color encoding and filtering design | Point-NN · S3DIS, 2-way 1-shot, S0 | mIoU: 42.12 -> 49.45 | Helped |
| add Seg-NN’s position-and-color encoding and filtering design | Point-NN · ScanNet, 2-way 1-shot, S0 | mIoU: 28.85 -> 36.80 | Helped |
| add QUEST (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 47.32 -> 64.84 | Helped |
| use self-correlation only instead of no QUEST (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 47.32 -> 49.05 | Helped |
| use cross-correlation only instead of no QUEST (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 47.32 -> 62.72 | Helped |
| add self-correlation to cross-correlation in QUEST (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 62.72 -> 64.84 | Helped |
| replace QUEST with a linear projection (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 64.84 -> 55.28 | Hurt |
| replace Seg-NN encoder with Point-NN backbone (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 64.84 -> 46.31 | Hurt |
| replace Seg-NN encoder with Point-PN backbone (ablation) | Seg-PN · S3DIS, 2-way 1-shot, S0 | mIoU: 64.84 -> 54.74 | Hurt |
Surprising findings
- Seg-NN without any training outperforms parametric DGCNN and ProtoNet baselines in several few-shot settings.
- Seg-NN shows only marginal performance difference between seen and unseen classes, whereas DGCNN generalizes much worse.
Limitations
- Seg-NN alone trails the best parametric methods on several settings (e.g., 49.53% vs PAP3D 59.45% on S3DIS 2-way 1-shot S0)
- Evaluated only on S3DIS and ScanNet scene datasets with cross-validated class splits
- Hyperparameters such as number of manipulation layers and frequency scale are tuned per setting (e.g., best layers differ between Seg-NN and Seg-PN)
Experimental setup
Datasets: S3DIS ScanNet ModelNet40 ScanObjectNN ShapeNetPart
Benchmarks: S3DIS ScanNet ModelNet40 ScanObjectNN ShapeNetPart
Metrics: mIoU training time parameter count inference speed accuracy
Baselines: Point-NN DGCNN ProtoNet MPTI AttMPTI BFG 2CBR PAP3D
Builds on: Point-NN DGCNN masked average pooling prototype networks PAP3D
Closest papers by meaning
- SegGraph: Leveraging Graphs of SAM Segments for Few-Shot 3D Part Segmentation (NeurIPS 2025 · similarity 0.87)
- APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · similarity 0.85)
- OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging (CVPR 2025 · similarity 0.85)
- Unlocking the Potential of Pre-trained Vision Transformers for Few-Shot Semantic Segmentation through Relationship Descriptors (CVPR 2024 · similarity 0.85)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.84)
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.