Research map / Detection and segmentation / 3D detection and driving
Superpoint Transformer for 3D Scene Instance Segmentation
Jiahao Sun, Chunmei Qing, Junpeng Tan, Xiangmin Xu
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPFormer's superpoint-query instance segmentation instead of the strongest listed baseline | ScanNetv2 hidden test · ScanNetv2 hidden test | mAP: 50.6 -> 54.9 | Improved |
| Use SPFormer rather than SSTNet | ScanNetv2 validation · ScanNetv2 validation | mAP: 49.4 -> 56.3 | Improved |
| Use SPFormer rather than the strongest listed baseline | S3DIS Area 5 · S3DIS Area 5 | AP50: 66.1 -> 66.8 | Improved |
| Use SPFormer rather than the strongest listed baseline | S3DIS 6-fold cross-validation · S3DIS 6-fold cross-validation | AP50: 68.9 -> 69.2 | Improved |
| Add superpoint pooling between the point backbone and query decoder (ablation) | SPFormer · ScanNetv2 validation | mAP: 34.3 -> 55.0 | Improved |
| Replace box matching with mask matching (ablation) | SPFormer with superpoint pooling · ScanNetv2 validation | mAP: 49.9 -> 55.0 | Improved |
| Add the IoU-aware score branch (ablation) | SPFormer with superpoint pooling and mask matching · ScanNetv2 validation | mAP: 55.0 -> 56.3 | Improved |
| Add iterative predictions from each decoder layer (ablation) | SPFormer transformer decoder · ScanNetv2 validation | mAP: 51.0 -> 52.5 | Improved |
| Add superpoint attention masks (ablation) | SPFormer with iterative prediction · ScanNetv2 validation | mAP: 52.5 -> 56.0 | Improved |
| Add position encoding to query vectors (ablation) | SPFormer with iterative prediction and attention masks · ScanNetv2 validation | mAP: 56.0 -> 55.6 | Hurt |
| Place cross-attention before self-attention (ablation) | SPFormer transformer decoder · ScanNetv2 validation | mAP: 56.0 -> 56.3 | Improved |
| Use 6 rather than 1, 3, or 12 transformer decoder layers (ablation) | SPFormer · ScanNetv2 validation | mAP: 49.1; 54.7; 56.3; 55.3 | Mixed |
Closest papers by meaning
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.89)
- SA3DIP: Segment Any 3D Instance with Potential 3D Priors (NeurIPS 2024 · similarity 0.88)
- Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking (CVPR 2025 · similarity 0.86)
- SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection (AAAI 2024 · similarity 0.86)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.86)
- Relation3D : Enhancing Relation Modeling for Point Cloud Instance Segmentation (CVPR 2025 · 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.
