Research map / Point clouds / Segmentation and detection
LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds
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 LogoSP, an unsupervised 3D semantic segmentation pipeline that distills DINOv2 features into 3D, grows superpoints, and groups superpoints via graph Fourier transform eigenvector patterns to produce semantic pseudo-labels.
Problem
Unsupervised semantic segmentation of raw 3D point clouds without human labels; existing methods rely on per-point local features with simple grouping, missing richer global semantic priors.
Approach
Builds a global superpoint graph over the whole dataset, computes its normalized Laplacian eigenvectors as global frequency-domain patterns, groups these patterns with K-means, and clusters superpoints by their pattern-membership vectors to obtain semantic pseudo-labels, combined with 2D-to-3D DINOv2 distillation and bottom-up superpoint growing.
Key results
- On ScanNet validation, LogoSP reaches 35.8% mIoU vs 29.6% for PointDC-DINOv2 and 25.4% for GrowSP.
- On ScanNet online hidden test, LogoSP achieves 32.7% mIoU vs 26.9% for GrowSP and 22.9% for PointDC.
- On S3DIS Area-5, LogoSP reaches 46.5% mIoU, slightly above GrowSP's 44.5%; on nuScenes it reaches 20.1% mIoU vs 18.2% for PointDC-DINOv2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LogoSP local-global superpoint grouping | Unsupervised 3D segmentation · ScanNet validation | mIoU (%): 29.6 -> 35.8 | Helped |
| Add LogoSP local-global superpoint grouping | Unsupervised 3D segmentation · ScanNet online test | mIoU (%): 26.9 -> 32.7 | Helped |
| Add LogoSP local-global superpoint grouping | Unsupervised 3D segmentation · S3DIS Area-5 | mIoU (%): 44.5 -> 46.5 | Helped |
| Add LogoSP local-global superpoint grouping | Unsupervised 3D segmentation · S3DIS 6-fold cross-validation | mIoU (%): 44.6 -> 46.3 | Helped |
| Add LogoSP local-global superpoint grouping | Unsupervised 3D segmentation · nuScenes validation | mIoU (%): 18.2 -> 20.1 | Helped |
| Add 2D-3D distillation (ablation) | LogoSP · ScanNet validation | mIoU (%): 26.8 -> 35.8 | Helped |
| Use DINO or STEGO instead of DINOv2 for feature distillation (ablation) | LogoSP · ScanNet validation | mIoU (%): 35.8 -> 25.1; 35.8 -> 25.8 | Hurt |
| Add bottom-up superpoint growing (ablation) | LogoSP · ScanNet validation | mIoU (%): 29.3 -> 35.8 | Helped |
| Vary final superpoint count Mᵀ (30, 40, 50, or 60) (ablation) | LogoSP · ScanNet validation | mIoU (%): 32.1; 34.7; 35.8; 32.4 | Mixed |
| Vary initial superpoint count M¹ (60, 80, or 100) (ablation) | LogoSP · ScanNet validation | mIoU (%): 35.7; 35.8; 32.1 | Mixed |
| Add top-down semantic pseudo-label generation (ablation) | LogoSP · ScanNet validation | mIoU (%): 31.4 -> 35.8 | Helped |
| Vary grouped global-pattern count S′ (30, 40, 50, 60, or 70) (ablation) | LogoSP · ScanNet validation | mIoU (%): 32.4; 32.6; 35.8; 35.6; 34.3 | Mixed |
Surprising findings
- On S3DIS, LogoSP is only about 2 mIoU points above GrowSP, attributed to sparse RGB images limiting 2D-to-3D distillation.
- On nuScenes, the paper notes performance is mixed, suggesting DINOv2 features struggle to generalize to self-driving scenes.
Limitations
- Gains on S3DIS are small because sparse 2D images yield less discriminative distilled 3D features.
- Performance on the outdoor nuScenes dataset is mixed, with DINOv2 features not generalizing well to driving scenes.
- Requires paired RGB/D images for 2D-to-3D distillation and assumes the number of classes C is known.
Experimental setup
Datasets: ScanNet S3DIS nuScenes
Benchmarks: ScanNet S3DIS nuScenes
Metrics: mIoU OA mAcc
Baselines: GrowSP PointDC PointDC-DINOv2 IIC PiCIE K-means
Builds on: GrowSP PointDC OpenScene DINOv2
Closest papers by meaning
- PDF: A Probability-Driven Framework for Open World 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.87)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.87)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.87)
- Point-SAM: Promptable 3D Segmentation Model for Point Clouds (ICLR 2025 · similarity 0.87)
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (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.