Research map / Point clouds / Segmentation and detection
Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic 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 TaylorSeg, a pretraining-free few-shot point cloud semantic segmentation network built on a Taylor-series-inspired local structure fitting convolution (TaylorConv) and an Adaptive Push-Pull prototype alignment module, in non-parametric and parametric variants.
Problem
Few-shot point cloud semantic segmentation of unseen categories suffers from costly DGCNN pretraining, class bias from pretraining on seen classes, and weak local structure representation in the backbone.
Approach
Model local point cloud structure as a polynomial (Taylor-series) fitting problem: a low-order term aggregates max-pooled neighbor features and a high-order term uses a kernel over relative neighbor features (generalizing affine and radial basis functions). A parameter-free variant avoids training entirely, and an APP module pushes and pulls support and query prototype features to reduce their dis
Key results
- Under 2-way 1-shot on S3DIS, TaylorSeg-PN reaches 69.12% average mIoU, +2.71 over Seg-PN (+2.28 on S0 split).
- On ScanNet 2-way 1-shot, TaylorSeg-PN achieves 69.14% average mIoU, +5.40 over the prior best (Seg-PN at 63.74%).
- TaylorSeg-NN, with zero parameters and no training, reaches 51.41% average mIoU on S3DIS 2-way 1-shot, +1.88 over Seg-NN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TaylorSeg-NN instead of Seg-NN | Few-shot point cloud segmentation · S3DIS 2-way 1-shot | mIoU: +1.88 | Helped |
| Use TaylorSeg-PN instead of Seg-PN | Few-shot point cloud segmentation · S3DIS 2-way 1-shot | mIoU: +2.71 | Helped |
| Use TaylorSeg-NN instead of Seg-NN | Few-shot point cloud segmentation · ScanNet 2-way 1-shot | mIoU: +1.31% | Helped |
| Use TaylorSeg-PN instead of the previous state-of-the-art method | Few-shot point cloud segmentation · ScanNet 2-way 1-shot | mIoU: +5.40% | Helped |
| Set s=1 and p learnable in HiConv rather than use ABF or RBF (ablation) | TaylorSeg-PN · S3DIS 2-way 1-shot | mIoU: 67.50 -> 69.12; 59.28 -> 69.12 | Helped |
| Set s=1 and p learnable in HiConv rather than set s=0 with p learnable (ablation) | TaylorSeg-PN · S3DIS 2-way 1-shot | mIoU: 66.97 -> 69.12 | Helped |
| Set s=1 and p learnable in HiConv rather than use ABF or RBF (ablation) | TaylorSeg-PN · S3DIS 3-way 1-shot | mIoU: 62.48 -> 71.34; 58.65 -> 71.34 | Helped |
| Set s=1 and p learnable in HiConv rather than set s=0 with p learnable (ablation) | TaylorSeg-PN · S3DIS 3-way 1-shot | mIoU: 62.22 -> 71.34 | Helped |
| Add ACP to TaylorSeg-PN without APP (ablation) | TaylorSeg-PN · S3DIS 2-way 1-shot | mIoU: 51.05 -> 53.13 | Helped |
| Add CAP to TaylorSeg-PN without APP (ablation) | TaylorSeg-PN · S3DIS 2-way 1-shot | mIoU: 51.05 -> 66.59 | Helped |
| Use both ACP and CAP rather than CAP alone (ablation) | TaylorSeg-PN · S3DIS 2-way 1-shot | mIoU: 66.59 -> 69.12 | Helped |
| Add ACP to TaylorSeg-PN without APP (ablation) | TaylorSeg-PN · S3DIS 3-way 1-shot | mIoU: 41.77 -> 43.10 | Helped |
Surprising findings
- Ablation shows that using only the ABF or RBF kernel in HiConv gives lower performance than the proposed high-order kernel (s=1, p learnable), and the RBF variant alone drops to 59.28% average in 2-way 1-shot.
Limitations
- Authors note TaylorConv has high computational complexity that they plan to reduce.
- Experiments are limited to S3DIS and ScanNet and to 2/3-way settings.
Experimental setup
Datasets: S3DIS ScanNet
Benchmarks: S3DIS ScanNet
Metrics: mIoU
Baselines: Point-NN Seg-NN DGCNN ProtoNet MPTI AttMPTI BFG 2CBR PAP3D Seg-PN
Builds on: Point-NN Seg-NN AttMPTI PointNet++ DGCNN
Closest papers by meaning
- DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation (ICML 2025 · similarity 0.93)
- EPSegFZ: Efficient Point Cloud Semantic Segmentation for Few- and Zero-Shot Scenarios with Language Guidance (AAAI 2026 · similarity 0.90)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model (CVPR 2025 · similarity 0.87)
- Generated and Pseudo Content guided Prototype Refinement for Few-shot Point Cloud Segmentation (NeurIPS 2024 · similarity 0.87)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.86)
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.