Research map / Point clouds / Segmentation and detection
Spherical Mask: Coarse-to-Fine 3D Point Cloud Instance Segmentation with Spherical Representation
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 Spherical Mask, a coarse-to-fine 3D point cloud instance segmentation method that represents instances as spherical 3D polygons and virtually migrates points radially to refine masks, reaching state of the art on ScanNetV2, S3DIS, and STPLS3D.
Problem
Coarse-to-fine 3D instance segmentation lags behind grouping, kernel and transformer methods because AABB coarse detection overestimates instance size and false negatives from the coarse stage propagate into refinement.
Approach
Estimate each instance as a 3D polygon in spherical coordinates (center plus per-sector radial distances), then refine all foreground points, including false negatives outside the polygon, by learning a single radial offset per point, supervised by margin-based misclassification and sector cohesion losses.
Key results
- On ScanNetV2 validation, Spherical Mask reaches 62.3 mAP and 79.9 AP50, versus 58.4 mAP for MAFT.
- On ScanNetV2 hidden test, it achieves 61.6 mAP, the highest among compared methods.
- On S3DIS Area 5 it reports 60.5 mAP, and on STPLS3D validation 52.2 mAP versus 49.2 for ISBNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Radial Point Migration (RPM) (ablation) | RID · ScanNetV2 validation | mAP: 51.6 -> 62.3 | Helped |
| add misclassification correction loss Lmc (ablation) | RID · ScanNetV2 validation | mAP: 51.6 -> 58.5 | Helped |
| add sector cohesion loss Lsc (ablation) | RID · ScanNetV2 validation | mAP: 51.6 -> 56.5 | Helped |
| combine Lmc and Lsc (ablation) | Spherical Mask · ScanNetV2 validation | mAP: 58.5 -> 62.3 | Helped |
| increase Nθ/Nφ from 4/4 to 5/5 (ablation) | Spherical Mask · ScanNetV2 validation | mAP: 61.2 -> 62.3 | Helped |
| increase Nθ/Nφ from 5/5 to 6/6 (ablation) | Spherical Mask · ScanNetV2 validation | mAP: 62.3 -> 60.6 | Hurt |
| increase seed count from 1024 to 2048 (ablation) | Spherical Mask · ScanNetV2 validation | mAP: 62.3 -> 61.8 | Hurt |
| vary seed and vote counts across tested settings (ablation) | Spherical Mask · ScanNetV2 validation | mAP: ±0.6 | No clear effect |
Surprising findings
- Increasing the sector grid from 5/5 to 6/6 reduced mAP by 2.8 points, which the authors attribute to many empty sectors yielding biased ray targets.
Limitations
- Performance is sensitive to the number of angular sectors (Ntheta, Nphi), with too many sectors degrading results.
- Superpoint aggregation is used on ScanNetV2 to align final masks, adding a dependency beyond the core method.
Experimental setup
Datasets: ScanNetV2 S3DIS STPLS3D
Benchmarks: ScanNetV2 S3DIS Area 5 STPLS3D
Metrics: mAP AP50 AP25 mPrec50 mRec50
Baselines: 3D-SIS GSPN TD3D PointGroup SSTNet SoftGroup PBNet DKNet ISBNet Mask3D MAFT QueryFormer
Builds on: ISBNet PolarMask 3D-BoNet DyCo3D
Closest papers by meaning
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.89)
- Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking (CVPR 2025 · similarity 0.88)
- Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance (CVPR 2024 · similarity 0.87)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.86)
- MaskClustering: View Consensus based Mask Graph Clustering for Open-Vocabulary 3D Instance Segmentation (CVPR 2024 · similarity 0.86)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 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.