Research map / Point clouds / Segmentation and detection
Hyperbolic Uncertainty-Aware Few-Shot Incremental 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 HIPO, a hyperbolic (Poincaré ball) few-shot class-incremental 3D point cloud segmentation method using CLIP-guided ideal prototypes and an uncertainty-aware Busemann loss, evaluated on S3DIS and ScanNetv2.
Problem
Few-shot class-incremental learning for 3D point cloud semantic segmentation, where new classes arrive with few labeled samples and models suffer overfitting and catastrophic forgetting in Euclidean embeddings.
Approach
Embed features on a Poincaré hyperbolic ball and pre-assign orthogonal ideal prototypes on the boundary, positioned using CLIP text-embedding semantic ranking; align class features to their prototypes via a Busemann loss with a hyperbolic-distance uncertainty penalty, without rehearsal.
Key results
- On S3DIS 10-3T (5-shot), HIPO reaches 33.40% Last mIoU versus 25.02% for C-FSCIL and 20.02% for LGKD.
- HIPO averages 31.13% Last mIoU across S3DIS settings, versus 26.50% for 3DPC-CISS (the prior SOTA), and reduces average forgetting to 14.16%.
- With a DGCNN backbone on S3DIS, HIPO achieves 29.90% Last mIoU, above BiDist's 27.70%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · S3DIS, average of 4-3T, 7-3T, 10-3T, 5-shot | Last mIoU (%): 26.50 -> 31.13 | Helped |
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · S3DIS, average of 4-3T, 7-3T, 10-3T, 5-shot | AIA (%): 37.73 -> 40.30 | Helped |
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · S3DIS, average of 4-3T, 7-3T, 10-3T, 5-shot | F(T)_Last (%): 17.46 -> 14.16 | Helped |
| Use HIPO instead of C-FSCIL | PointTransformer FSCIL · ScanNetv2, average of 10-2T, 10-5T, 15-5T, 5-shot | Last mIoU (%): 9.25 -> 11.68 | Helped |
| Use HIPO instead of LGKD | PointTransformer FSCIL · ScanNetv2, average of 10-2T, 10-5T, 15-5T, 5-shot | AIA (%): 21.65 -> 22.27 | Helped |
| Use HIPO instead of C-FSCIL | PointTransformer FSCIL · ScanNetv2, average of 10-2T, 10-5T, 15-5T, 5-shot | F(T)_Last (%): 18.11 -> 18.12 | No clear effect |
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · ScanNetv2 → S3DIS, average of 10-4T and 10-8T, 5-shot | Last mIoU (%): 19.95 -> 20.60 | Helped |
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · ScanNetv2 → S3DIS, average of 10-4T and 10-8T, 5-shot | AIA (%): 29.16 -> 29.81 | Helped |
| Use HIPO instead of 3DPC-CISS | PointTransformer FSCIL · ScanNetv2 → S3DIS, average of 10-4T and 10-8T, 5-shot | F(T)_Last (%): 16.01 -> 16.05 | No clear effect |
| Vary uncertainty penalty strength ϕ from 0.00 to 1.00 (ablation) | HIPO · S3DIS and ScanNetv2, 5-shot; 4-3T, 7-3T, 10-3T, 10-2T, 10-5T, 15-5T | AIA (%): S3DIS 4-3T: 42.87, 43.40, 43.27, 43.87, 43.22; 7-3T: 36.13, 36.46, 36.80, 36.70, 36.06; 10-3T: 39.95, 40.25, 40.10, 40.0 | Mixed |
| Vary uncertainty penalty strength ϕ from 0.00 to 1.00 (ablation) | HIPO · S3DIS and ScanNetv2, 5-shot; 4-3T, 7-3T, 10-3T, 10-2T, 10-5T, 15-5T | F(T)_Last (%): S3DIS 4-3T: 17.70, 17.86, 18.23, 17.70, 18.96; 7-3T: 11.60, 11.20, 11.10, 11.10, 12.10; 10-3T: 14.30, 13.70, 14.00, 14.0 | Mixed |
| Use near-zero curvature c = 0.0001 instead of c = −1.00 (ablation) | HIPO · S3DIS and ScanNetv2, 5-shot; 4-3T, 7-3T, 10-3T, 10-2T, 10-5T, 15-5T | AIA (%): S3DIS: 42.95, 36.23, 40.20; ScanNetv2: 16.54, 27.16, 24.40 | Mixed |
Surprising findings
- Euclidean-like curvature (c=0.0001, 0.001) underperforms the negative-curvature hyperbolic setting in most settings.
- Without the uncertainty penalty (phi=0), overconfident predictions and lower AIA result, while too high a penalty increases inter-class confusion.
Limitations
- Does not handle outliers or unknown classes appearing without reference across incremental sessions (stated as future work).
- Cross-dataset gains are modest (about +0.5% Last and AIA on ScanNetv2 to S3DIS).
Experimental setup
Datasets: S3DIS ScanNetv2
Benchmarks: S3DIS ScanNetv2
Metrics: mIoU Average Incremental Accuracy Last accuracy Average Forgetting Rate
Baselines: LwF C-FSCIL LGKD SDCOT 3DPC-CISS BiDist CLOM TEEN FACT ALICE
Builds on: Ideal prototypes (Ghadimi et al.) Busemann loss PointTransformer CLIP LwF C-FSCIL
Closest papers by meaning
- Generalized Zero-Shot Learning for Point Cloud Segmentation with Evidence-Based Dynamic Calibration (AAAI 2025 · similarity 0.89)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.89)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (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)
- 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.