Research map / Point clouds / Point cloud classification
X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition
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 X-3D, which builds explicit PointHop-style local structure descriptors in input 3D space and uses them to generate shared dynamic kernels for point cloud segmentation, classification, part segmentation and detection, plugged into several existing backbones.
Problem
Implicit high-dimensional structure modeling of point cloud local neighborhoods (relation vectors plus dynamic kernels) loses explicit geometric structure, creating a gap between embedding-space local structure and the original input geometry.
Approach
Instead of generating kernels per relation vector, generate one shared dynamic kernel per neighborhood from an explicit 3D local structure descriptor (PointHop-style octant centroids), with adaptive denoising via cross-attention and geometrically limited neighborhood context propagation.
Key results
- X-3D on DeLA reaches 90.7% OA on ScanObjectNN and 79.2% mIoU on S3DIS 6-fold, 74.3% on Area 5, as reported in the abstract and tables.
- Adding X-3D to PointNet++ improves S3DIS 6-fold mIoU by 7.4 points (54.5 to 62.9).
- On GroupFree with X-3D, ScanNetV2 detection reaches 69.0 mAP@0.25 and 51.1 mAP@0.50, and SUN RGB-D reaches 64.5 and 46.9, with about 0.8 GFLOPs extra cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add X-3D to DeLA | DeLA · S3DIS 6-fold | mIoU: 78.3 -> 79.2 | Helped |
| Add X-3D to DeLA | DeLA · S3DIS Area-5 | mIoU: 74.1 -> 74.3 | Helped |
| Add X-3D to DeLA | DeLA · ScanNet validation | mIoU: 75.9 -> 76.3 | Helped |
| Add X-3D | GroupFree6,256 · ScanNetV2 | mAP@0.50: 48.9 -> 51.1 | Helped |
| Add X-3D | GroupFree6,256 · SUN RGB-D | mAP@0.50: 45.2 -> 46.9 | Helped |
| Add X-3D to DeLA | DeLA · ScanObjectNN | OA: 90.4 -> 90.7 | Helped |
| Add X-3D relational position embedding | Point Transformer · ScanNet | mIoU: 70.6 -> 73.9 | Helped |
| Use a shared structure kernel instead of concatenating explicit structure (ablation) | PointMetaBase-S · ScanObjectNN | OA: 88.3 -> 88.8 | Helped |
| Generate separate vector kernels instead of one shared structure kernel (ablation) | PointMetaBase-S · ScanObjectNN | OA: 88.8 -> 88.6 | Hurt |
| Add denoising (ablation) | PointMetaBase · S3DIS Area5 | mIoU: 70.9 -> 71.3 | Helped |
| Add neighborhood context propagation (ablation) | PointMetaBase · S3DIS Area5 | mIoU: 70.9 -> 71.4 | Helped |
| Add neighborhood context propagation (ablation) | PointMetaBase · ScanObjectNN | OA: 88.5 -> 88.6 | No clear effect |
Surprising findings
- Generating a shared structure kernel per neighborhood gives better OA (88.8) than per-vector kernels (88.6) on ScanObjectNN while costing far fewer GFLOPs (0.59 vs 0.83).
Limitations
- Explicit structure descriptor (PointHop octant centroids) was chosen empirically among a few options; the property-based design is heuristic.
- Gains on some backbones are small (e.g., +0.2 to +0.5 mIoU on several S3DIS and ScanNet settings).
Experimental setup
Datasets: S3DIS ScanNetV2 ScanObjectNN SUN RGB-D ShapeNetPart KITTI
Benchmarks: S3DIS ScanNetV2 ScanObjectNN SUN RGB-D ShapeNetPart KITTI
Metrics: mIoU mAcc OA mAP@0.25 mAP@0.50 GFLOPs #Params
Baselines: PointNet++ PointMetaBase DeLA Point Transformer RepSurf KPConv PointNeXt GroupFree 3DSSD
Builds on: PointMetaBase PointNet++ PointHop Point Transformer GroupFree
Closest papers by meaning
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.89)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.88)
- AGILE3D: Attention Guided Interactive Multi-object 3D Segmentation (ICLR 2024 · similarity 0.87)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.87)
- Point-SAM: Promptable 3D Segmentation Model for Point Clouds (ICLR 2025 · similarity 0.87)
- Novel Class Discovery for Point Cloud Segmentation via Joint Learning of Causal Representation and Reasoning (NeurIPS 2025 · 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.