Research map / Detection and segmentation / 3D detection and driving
SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection
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 SPGroup3D, an anchor-free one-stage indoor 3D detector that groups unsupervised superpoints via superpoint attention and superpoint-voxel fusion, with geometry-aware voting and multiple matching, reaching SOTA one-stage mAP on ScanNet V2, SUN RGB-D and S3DIS.
Problem
Indoor 3D object detection with bottom-up voting-and-grouping relies on instance-agnostic groupings (e.g., ball query), producing mixed-instance proposals and inaccurate regression; conventional voting also hurts centerness-based filtering in anchor-free detectors.
Approach
Partition point clouds into semantically consistent, spatially similar superpoints and use them as instance-aware proposal groups; geometry-aware voting merges seed and vote voxels to keep relative geometry to object centers; superpoint attention (kNN in coordinate and feature space) and superpoint-voxel sparse-conv fusion iteratively refine features; multiple matching with a DETR-style cost assig
Key results
- Achieves 74.3 mAP@0.25 and 59.6 mAP@0.5 on ScanNet V2 (best-of-runs), +1.1 over prior one-stage SOTA on mAP@0.25.
- Reaches 65.4 mAP@0.25 on SUN RGB-D (+1.2) and 69.2 mAP@0.25 on S3DIS (+2.5) on the validation sets.
- Ablation on ScanNet V2: geometry-aware voting, superpoint grouping and multiple matching raise mAP@0.25 from 68.2 to 73.5 over the VoteNet-style baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add geometry-aware voting (ablation) | SPGroup3D base model · ScanNet V2 val | mAP@0.25: 68.2 to 71.3 | Helped |
| Add geometry-aware voting (ablation) | SPGroup3D base model · ScanNet V2 val | mAP@0.5: 53.2 to 57.0 | Helped |
| Add superpoint-based grouping (ablation) | SPGroup3D with geometry-aware voting · ScanNet V2 val | mAP@0.25: 71.3 to 72.7 | Helped |
| Add superpoint-based grouping (ablation) | SPGroup3D with geometry-aware voting · ScanNet V2 val | mAP@0.5: 57.0 to 57.9 | Helped |
| Add multiple matching (ablation) | SPGroup3D with geometry-aware voting and grouping · ScanNet V2 val | mAP@0.25: 72.7 to 73.5 | Helped |
| Add multiple matching (ablation) | SPGroup3D with geometry-aware voting and grouping · ScanNet V2 val | mAP@0.5: 57.9 to 58.3 | Helped |
| Use superpoint-based grouping rather than semantic-agnostic grouping (ablation) | SPGroup3D grouping strategy · ScanNet V2 val | mAP@0.25: 68.2 -> 73.5 | Helped |
| Use superpoint-based grouping rather than semantic-agnostic grouping (ablation) | SPGroup3D grouping strategy · ScanNet V2 val | mAP@0.5: 53.2 -> 58.3 | Helped |
| Preserve both pre- and post-voting features instead of only pre-voting features (ablation) | SPGroup3D geometry-aware voting · ScanNet V2 val | mAP@0.25: 72.0 to 73.5 | Helped |
| Preserve both pre- and post-voting features instead of only post-voting features (ablation) | SPGroup3D geometry-aware voting · ScanNet V2 val | mAP@0.5: 56.2 to 58.3 | Helped |
| Replace MLP with sparse convolution in superpoint-voxel fusion (ablation) | SPGroup3D · ScanNet V2 val | mAP@0.25: 72.2 -> 73.5 | Helped |
| Replace MLP with sparse convolution in superpoint-voxel fusion (ablation) | SPGroup3D · ScanNet V2 val | mAP@0.5: 57.4 -> 58.3 | Helped |
Surprising findings
- Using only the features before voting (72.0 mAP@0.25) beats using only features after voting (70.9), but combining both gives 73.5.
- On SUN RGB-D, SPGroup3D does not achieve the best mAP@0.5 (47.1 vs FCAF3D 48.9).
Limitations
- Inference is slower than some prior one-stage baselines such as VoteNet and FCAF3D.
- Lower mAP@0.5 than FCAF3D on SUN RGB-D, and weaker on some classes like sofa and bookshelf in the 0.25 setting.
Experimental setup
Datasets: ScanNet V2 SUN RGB-D S3DIS
Benchmarks: ScanNet V2 SUN RGB-D S3DIS
Metrics: mAP@0.25 mAP@0.5
Baselines: VoteNet 3D-MPA HGNet MLCVNet GSDN H3DNet BRNet 3DETR VENet GroupFree RBGNet HyperDet3D
Builds on: VoteNet FCAF3D CAGroup3D DETR BiResNet
Closest papers by meaning
- Cubify Anything: Scaling Indoor 3D Object Detection (CVPR 2025 · similarity 0.88)
- Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping (NeurIPS 2025 · similarity 0.87)
- Geometry-Aware 3D Salient Object Detection Network (AAAI 2025 · similarity 0.87)
- UniDet3D: Multi-dataset Indoor 3D Object Detection (AAAI 2025 · similarity 0.87)
- LION: Linear Group RNN for 3D Object Detection in Point Clouds (NeurIPS 2024 · similarity 0.86)
- Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection (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.