Research map / Point clouds / Point cloud classification
Geometric Feature Embedding for Effective 3D Few-Shot Class Incremental Learning
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 3D-FLEG, which augments frozen 3D-VLM text features with geometric features from spectral-clustering/Laplacian-eigenmap projection clusters fused via a small Transformer, for 3D few-shot class incremental learning.
Problem
3D few-shot class incremental learning must learn new point cloud classes from few real-scanned samples without forgetting old classes, while existing 3D foundation-model methods depend heavily on hand-crafted prompts and elaborate training strategies.
Approach
Reduce prompt dependence by injecting category-related geometric features, extracted from point cloud features via spectral-clustering-based dynamic projection clusters with learnable basis weights, into text features through a Transformer-based geometric embedding module, while keeping the pre-trained encoders frozen.
Key results
- Within-dataset ModelNet 3D-FLEG reaches 98.3 at session 0 and 87.3 at session 4 with a drop of 11.2%, versus 22.5% for C3PR.
- ShapeNet to ScanObjectNN: 3D-FLEG reaches 93.8 at session 0 and 86.8 at session 3, with a relative drop of 7.5%.
- Ablation on ModelNet to ScanObjectNN shows dynamic projection clusters plus attention weights raise final-session accuracy from 84.1 to 86.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · ModelNet, 40 classes | average accuracy: 70.9 -> 87.3 | Helped |
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · CO3D, 50 classes | average accuracy: 53.8 -> 57.3 | Helped |
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · ShapeNet, 55 classes | average accuracy: 74.7 -> 83.0 | Helped |
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · ShapeNet → CO3D, 89 classes | average accuracy: 72.6 -> 76.8 | Helped |
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · ModelNet → ScanObjectNN, 37 classes | average accuracy: 79.2 -> 86.8 | Helped |
| add geometric feature extraction and embedding to the 3D FSCIL method | 3D-FLEG · ShapeNet → ScanObjectNN, 59 classes | average accuracy: 85.6 -> 86.3 | Helped |
| add dynamic geometric feature projection clusters (ablation) | 3D-FLEG · ModelNet → ScanObjectNN, stage 3 | average accuracy: 84.1 -> 84.9 | Helped |
| add attention weights to dynamic geometric feature projection clusters (ablation) | 3D-FLEG · ModelNet → ScanObjectNN, stage 3 | average accuracy: 84.9 -> 86.8 | Helped |
| add dynamic geometric feature projection clusters and attention weights (ablation) | 3D-FLEG · ModelNet → ScanObjectNN, stage 3 | harmonic accuracy: 76.2 -> 77.5 | Helped |
| use the Transformer-based geometric embedding module instead of LSTM or FCNN (ablation) | 3D-FLEG · ModelNet → ScanObjectNN | feature embedding performance | Helped |
| use spectral clustering and Laplacian eigenmaps to construct projection clusters (ablation) | 3D-FLEG · ModelNet → ScanObjectNN | 3D FSCIL performance | Helped |
| use 256 basis vectors instead of 1024 (ablation) | 3D-FLEG · ModelNet → ScanObjectNN, stage 3 | average accuracy: 86.8 -> 85.0 | Hurt |
Surprising findings
- A simpler single prompt template outperforms more complex GPT-generated or 64-template prompts for the 3D foundation-model FSCIL pipeline, by roughly 6% in the final phase.
Limitations
- Relies on frozen pre-trained 3D foundation model features and its base-session training on synthetic data.
- Basis vector count is tied to the 1024-dim feature space, so generalizing to other feature dimensions requires extra compression or expansion modules.
Experimental setup
Datasets: ModelNet ShapeNet ScanObjectNN CO3D
Benchmarks: ModelNet ShapeNet ScanObjectNN CO3D
Metrics: average accuracy relative accuracy drop rate harmonic accuracy
Baselines: FT Joint LwF IL2M ScaIL EEIL FACT Sem-aware Microshape C3PR FoundationModel
Builds on: Uni3D FoundationModel (Ahmadi et al., 2024) Microshape C3PR
Closest papers by meaning
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.87)
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.85)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.85)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
- Activating Sparse Part Concepts for 3D Class Incremental Learning (CVPR 2025 · similarity 0.84)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · 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.