Research map / Point clouds / Shape and surface completion
Orthogonal Dictionary Guided Shape Completion Network for Point Cloud
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 ODGNet, a point cloud shape completion network combining a Seed Generation U-Net with learnable orthogonal shape dictionaries that inject shape priors into seed features, achieving state-of-the-art CD on PCN, ShapeNet-55/34 and KITTI.
Problem
Point cloud shape completion is ill-posed because encoded features from incomplete inputs lose information about missing regions, so coarse seed points lack detailed geometry and bottleneck the upsampling stage.
Approach
Learnable shape dictionaries trained with an orthogonality constraint store shape priors in feature space; coarse point features attend to the dictionary via cross-attention to be refined, and a Seed Generation U-Net with multi-level feature concatenation and seed features enhances seed representation.
Key results
- Achieves average CD-L1 of 6.50e-3 on PCN, improving over SeedFormer (6.74e-3) and AdaPoinTr (6.53e-3 reported; ODGNet best overall).
- Achieves average CD-L2 of 0.83e-3 on ShapeNet-55, but does not beat AdaPoinTr (0.81) and ProxyFormer/AdaPoinTr on F-score; F-score 0.437 is lower than several baselines.
- Dictionary Guidance Module lowers CD-L1 from 6.62e-3 (without) to 6.50e-3 (with orthogonal constraints) on PCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ODGNet instead of AdaPoinTr | PCN point cloud completion · PCN test | CD-L1 ×10−3: 6.53 -> 6.50 | Helped |
| Use ODGNet instead of AdaPoinTr | ShapeNet-55 point cloud completion · ShapeNet-55 test | CD-Avg ×10−3: 0.81 -> 0.83 | Hurt |
| Use ODGNet instead of AdaPoinTr | ShapeNet-55 point cloud completion · ShapeNet-55 test | F-Score: 0.503 -> 0.437 | Hurt |
| Use ODGNet instead of AdaPoinTr | ShapeNet-34 point cloud completion · Seen ShapeNet-34 test | CD-Avg ×10−3: 0.73 -> 0.75 | Hurt |
| Use ODGNet instead of AdaPoinTr | ShapeNet-34 point cloud completion · Unseen ShapeNet-21 test | CD-Avg ×10−3: 1.23 -> 1.29 | Hurt |
| Use ODGNet instead of SeedFormer | KITTI point cloud completion · KITTI | MMD (×10−3): 0.516 -> 0.349 | Helped |
| Replace FoldingNet seed generation with ODGNet Seed Generation U-Net (ablation) | Folding upsampling on PCN · PCN test | CD-L1 (×10−3): 14.31 -> 7.59 | Helped |
| Replace SnowFlake seed generation with ODGNet Seed Generation U-Net (ablation) | PSCU upsampling on PCN · PCN test | CD-L1 (×10−3): 7.04 -> 6.80 | Helped |
| Replace SeedFormer seed generation with ODGNet Seed Generation U-Net (ablation) | UpTrans upsampling on PCN · PCN test | CD-L1 (×10−3): 6.74 -> 6.50 | Helped |
| Add Dictionary Guidance Module (ablation) | ODGNet on PCN · PCN test | CD-L1 (×10−3): 6.62 -> 6.50 | Helped |
| Add orthogonal constraints to the learnable dictionaries (ablation) | ODGNet on PCN · PCN test | CD-L1 (×10−3): 6.55 -> 6.50 | Helped |
Surprising findings
- On PCN, the reported average CD-L1 of AdaPoinTr (6.53) is slightly worse than ODGNet's 6.50, but on ShapeNet-55 the paper's own CD-Avg (0.83) is worse than AdaPoinTr (0.81) and its F-score (0.437) trails several baselines.
Limitations
- No ground truth for KITTI, so Fidelity and MMD are not accurate measures of quality, as the authors note.
- Improvements over prior SOTA are small (e.g., 0.24e-3 CD-L1 on PCN), and some metrics on ShapeNet-55 are not best.
Experimental setup
Datasets: PCN ShapeNet-55 ShapeNet-34 KITTI
Benchmarks: PCN ShapeNet-55 ShapeNet-34 ShapeNet-21 KITTI
Metrics: CD-L1 CD-L2 F1-Score@1% Fidelity MMD
Baselines: FoldingNet TopNet AtlasNet PCN GR-Net PMP PoinTr SnowFlake FBNet ProxyFormer AdaPoinTr SeedFormer
Builds on: SeedFormer PoinTr TopNet DETR U-Net
Closest papers by meaning
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.88)
- SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance (AAAI 2025 · similarity 0.88)
- Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.87)
- PointAttN: You Only Need Attention for Point Cloud Completion (AAAI 2024 · similarity 0.87)
- FSC: Few-point Shape Completion (CVPR 2024 · similarity 0.87)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 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.