Research map / Point clouds / Segmentation and detection
Insightful Instance Features for 3D Instance 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 IKNE, which adds instance-wise knowledge aggregation (correlation regularization among candidates of the same instance) and truncated-SVD-based structural guidance to kernel-based and transformer-based 3D instance segmentation models.
Problem
Hundreds of instance candidates in 3D instance segmentation produce fragmented masks and imprecise spatial extents due to sparse, unordered point cloud noise, and existing methods underuse these candidates.
Approach
Softly aggregates scattered cues from highly correlated instance candidates via a correlation regularization loss, and guides candidate features with structural cues from a truncated-SVD low-rank reconstruction of the same features.
Key results
- On ScanNetV2 validation, IKNE reaches 62.9 mAP / 81.8 mAP50, above Spherical Mask (62.3 / 79.9).
- On S3DIS Area 5, IKNE reaches 61.1 mAP / 73.0 mAP50, and 65.8 / 76.9 in 6-fold cross-validation.
- On ScanNet200 validation, IKNE reaches 29.9 mAP versus 29.2 for MAFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IKA to the kernel-based method (ablation) | IKNE kernel-based pipeline · ScanNetV2 Val and S3DIS Area 5 | mAP: 56.8 -> 62.7; 56.3 -> 60.4 | Helped |
| Add ISG to the kernel-based method (ablation) | IKNE kernel-based pipeline · ScanNetV2 Val and S3DIS Area 5 | mAP: 56.8 -> 62.8; 56.3 -> 60.6 | Helped |
| Add IKA to the transformer-based method (ablation) | IKNE transformer-based pipeline · ScanNetV2 Val and S3DIS Area 5 | mAP: 58.4 -> 59.1; 56.6 -> 58.8 | Helped |
| Add ISG to the transformer-based method (ablation) | IKNE transformer-based pipeline · ScanNetV2 Val and S3DIS Area 5 | mAP: 58.4 -> 59.4; 56.6 -> 58.4 | Helped |
| Add IKA (ablation) | Kernel-based and transformer-based pipelines · ScanNetV2 Val | Average IoU: 0.8627 -> 0.8828; 0.8486 -> 0.8728 | Helped |
| Add ISG (ablation) | Kernel-based and transformer-based pipelines · ScanNetV2 Val | Average IoU: 0.8627 -> 0.8825; 0.8486 -> 0.8730 | Helped |
| Set correlation threshold τ to 0.6 (ablation) | Kernel-based pipeline · ScanNetV2 Val | mAP: 56.8 -> 53.6 | Hurt |
| Set correlation threshold τ to 0.9 (ablation) | Kernel-based pipeline · ScanNetV2 Val | mAP: 56.8 -> 62.9 | Helped |
| Vary the retained top-t percentage in truncated SVD (ablation) | IKNE kernel-based and transformer-based pipelines · ScanNetV2 and S3DIS | mAP | Mixed |
Limitations
- Performance is sensitive to the correlation threshold tau; values below 0.7 degrade results.
- Truncation ratio t must be tuned (chosen as 0.5); too low loses information and too high gives little noise reduction.
Experimental setup
Datasets: ScanNetV2 ScanNet200 S3DIS STPLS3D
Benchmarks: ScanNetV2 ScanNet200 S3DIS STPLS3D
Metrics: mAP mAP50 mAP25 IoU mPrec mRec
Baselines: ISBNet MAFT Mask3D Spherical Mask EASE Q-Former SoftGroup SPFormer PointGroup DyCo3D
Builds on: ISBNet MAFT Mask3D SoftGroup
Closest papers by meaning
- Edge-Aware 3D Instance Segmentation Network with Intelligent Semantic Prior (CVPR 2024 · similarity 0.89)
- SAI3D: Segment Any Instance in 3D Scenes (CVPR 2024 · similarity 0.89)
- Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance (CVPR 2024 · similarity 0.87)
- Relation3D : Enhancing Relation Modeling for Point Cloud Instance Segmentation (CVPR 2025 · similarity 0.87)
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · similarity 0.87)
- Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation (NeurIPS 2026 · 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.