Research map / Point clouds / Point cloud classification
Point Deformable Network with Enhanced Normal Embedding for Point Cloud Analysis
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 PDNet, an MLP-based point cloud network with a Point Deformable Aggregation Module (learned deformable reference points for long-range aggregation) and Enhanced Normal Embedding from least-squares normals.
Problem
MLP-based point cloud networks aggregate features only within local point groups (kNN or ball query) and fail to model long-range dependencies directly.
Approach
Generate deformable reference points initialized by FPS over all points, shift them with offsets and modulation scalars learned from whole-cloud point features, and aggregate their interpolated features into each query point; additionally embed least-squares estimated normals via an MLP (ENE) to strengthen per-point geometry.
Key results
- PDNet achieves 88.5% OA and 86.8% mAcc on ScanObjectNN PB_T50_RS, outperforming PointMetaBase (88.2% OA).
- PDNet-XXL reaches 78.3% mIoU on S3DIS 6-Fold and 72.3% mIoU on S3DIS Area-5.
- PDNet achieves 85.4% class mIoU and 87.2% instance mIoU on ShapeNetPart.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PDAM to PLAM (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 69.5±0.3 -> 70.4±0.3 | Helped |
| Add normal embedding to PLAM and PDAM (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.4±0.3 -> 70.6±0.2 | Helped |
| Use PDAM rather than PLAM alone in the deep stages (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 69.5±0.3 -> 70.1±0.2 | Helped |
| Add PDAM in Stage 4 (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 69.7 -> 70.1 | Helped |
| Apply PDAM in Stages 3 and 4 (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.1 -> 70.8 | Helped |
| Add PDAM in the early Stage 2 (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.8 -> 69.9 | Hurt |
| Initialize reference points with FPS instead of random positions (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.1 -> 70.8 | Helped |
| Initialize each reference point independently with FPS instead of placing all at the center (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.6 -> 70.8 | Helped |
| Apply PLAM and PDAM in parallel rather than successively (ablation) | PDNet-L · S3DIS Area-5 | mIoU (%): 70.4 -> 70.8 | Helped |
| Use PDNet-XXL instead of PointMetaBase-XXL | PointMetaBase-XXL · S3DIS 6-Fold | mIoU (%): +1.3 | Helped |
| Use PDNet-XXL instead of PointMetaBase-XXL | PointMetaBase-XXL · S3DIS Area-5 | mIoU (%): 71.3±0.7 -> 72.3 | Helped |
| Use PDNet instead of PointMetaBase | PointMetaBase · ShapeNetPart | i. mIoU: 87.1±0.0 -> 87.2 | Helped |
Surprising findings
- Applying PDAM in early stages decreases mIoU on S3DIS Area-5, so PDAM is beneficial mainly in deeper stages.
Limitations
- Gains over PointMetaBase on ScanObjectNN and ShapeNetPart are small (about 0.3 to 0.1 points).
- Normal estimation requires a kNN-based least-squares fit when normals are absent; ENE is not needed where normals exist.
- PDAM in early stages degrades performance, so the design is stage-dependent.
Experimental setup
Datasets: ScanObjectNN S3DIS ShapeNetPart
Benchmarks: ScanObjectNN PB_T50_RS S3DIS 6-Fold S3DIS Area-5 ShapeNetPart
Metrics: OA mAcc mIoU class mIoU instance mIoU
Baselines: PointNet PointNet++ PointCNN DGCNN PointMLP PointNeXt PointMetaBase GAM Point Transformer KPConv
Builds on: DCNv2 PointNeXt PointMetaBase PointNet++ KPConv
Closest papers by meaning
- Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework (ICLR 2022 · similarity 0.89)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.88)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 2025 · similarity 0.86)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.86)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.85)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.85)
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.