Research map / Point clouds / Segmentation and detection
DyPolySeg: Taylor Series-Inspired Dynamic Polynomial 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 DyPolySeg, a pre-training-free few-shot point cloud segmentation network that combines Taylor-inspired dynamic polynomial convolution, a Mamba block, and a prototype completion module.
Problem
Few-shot point cloud semantic segmentation suffers from domain shift and extra cost from pre-training, limited local geometric modeling of DGCNN backbones, and a categorical/structural gap between query and support prototypes.
Approach
A Dynamic Polynomial Convolution (DyPolyConv) that pairs low-order convolution with learned dynamic high-order polynomial terms (Taylor-inspired) for local geometry, stacked with Mamba blocks for global context, plus a lightweight Prototype Completion Module with self- and interactive enhancement to reduce support-query prototype bias.
Key results
- On S3DIS 2-way 1-shot, DyPolySeg achieves 72.92% mIoU vs. 66.41% for Seg-PN (+6.51).
- On ScanNet 2-way 1-shot, DyPolySeg achieves 71.89% mIoU vs. 63.74% for Seg-PN (+8.15).
- On ScanObjectNN (PB_T50_RS), DyPolySeg reports 90.8% accuracy with 2.64M parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Seg-PN with DyPolySeg | Few-shot point cloud segmentation · S3DIS, 2-way 1-shot; 2-way 5-shot; 3-way 1-shot; 3-way 5-shot | mIoU: 66.41 -> 72.92; 69.56 -> 75.66; 59.77 -> 66.24; 62.10 -> 67.92 | Helped |
| Replace Seg-PN with DyPolySeg | Few-shot point cloud segmentation · ScanNet, 2-way 1-shot; 2-way 5-shot; 3-way 1-shot; 3-way 5-shot | mIoU: 63.74 -> 71.89; 68.07 -> 72.46; 63.57 -> 69.45; 65.60 -> 69.18 | Helped |
| Replace PointMamba with DyPolySeg | ScanObjectNN PB T50 RS classification · ScanObjectNN PB T50 RS | accuracy: 89.3 -> 90.8 | Helped |
| Add DyHoConv to LoConv (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 48.68 -> 51.36 | Helped |
| Add Mamba Block (ablation) | DyPolySeg with LoConv and DyHoConv · S3DIS, 2-way 1-shot | mIoU: 51.36 -> 53.28 | Helped |
| Add PCM (ablation) | DyPolySeg with LoConv and DyHoConv · S3DIS, 2-way 1-shot | mIoU: 51.36 -> 70.48 | Helped |
| Add SEM and IEM to PCM (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 50.75 -> 72.92 | Helped |
| Increase the number of HoConvs from 1 to 8 (ablation) | DyPolySeg DyHoConv · S3DIS, 2-way 1-shot | mIoU: 69.33 -> 72.92 | Helped |
| Use 2 PCM layers instead of 0 (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 50.75 -> 72.92 | Helped |
| Increase PCM depth from 2 to 3 layers (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 72.92 -> 70.63 | Hurt |
| Remove IEM from PCM (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 70.59 -> 72.92 | Hurt |
| Remove SEM from PCM (ablation) | DyPolySeg · S3DIS, 2-way 1-shot | mIoU: 70.05 -> 72.92 | Hurt |
Surprising findings
- Stacking a third PCM layer slightly reduces performance (70.63% vs 72.92% in 2-way 1-shot), so two layers are optimal.
Limitations
- Relies on a 2048-point block protocol and k-NN with 16 neighbors, which may limit scalability to larger scenes.
- Gains on ScanObjectNN are reported without a clear description of the classification protocol in the main text.
Experimental setup
Datasets: S3DIS ScanNet ScanObjectNN
Benchmarks: S3DIS ScanNet ScanObjectNN
Metrics: mIoU accuracy
Baselines: DGCNN ProtoNet MPTI AttMPTI BFG 2CBR PAP3D Seg-PN
Builds on: TaylorSeg Seg-PN AttMPTI DGCNN Mamba
Closest papers by meaning
- Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · similarity 0.93)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Generated and Pseudo Content guided Prototype Refinement for Few-shot Point Cloud Segmentation (NeurIPS 2024 · 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 0.87)
- Collaborating Foundation Models for Domain Generalized Semantic Segmentation (CVPR 2024 · 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.