Research map / Point clouds / Shape and surface completion
CDPNet: Cross-Modal Dual Phases Network for Point Cloud Completion
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 CDPNet, a cross-modal dual-phase network that uses a single-view image for global shape structure and partial point clouds for local geometry, generating patch-based dense point clouds for point cloud completion on ShapeNet-ViPC.
Problem
Point cloud completion from partial scans suffers information loss in encoder-decoder pipelines because missing regions are invisible; the paper asks whether a single-view image can supply global shape information to guide completion.
Approach
Treats completion as cross-modal: a coarse point cloud reconstructed from the image is split into patches, and per-patch generators are guided by partial-point-cloud geometry and a style code from cross-attention fusion of image and point features (AdaIN-based style generator).
Key results
- CDPNet achieves average Chamfer Distance 1.066 (x10^3) on ShapeNet-ViPC, versus 1.367 for XMFNet and 1.550 for TopNet, the best among compared methods.
- Average F-Score of 0.869 for CDPNet versus 0.810 for XMFNet on ShapeNet-ViPC.
- Module ablation: removing all modules gives CD 1.850; dual phases plus multi-patch generator yields 1.103; adding feature fusion yields 1.066.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDPNet instead of XMFNet | Point cloud completion · ShapeNet-ViPC Avg | Chamfer Distance (×103): 1.367 -> 1.066 | Helped |
| Use CDPNet instead of XMFNet | Point cloud completion · ShapeNet-ViPC Avg | F-Score: 0.810 -> 0.869 | Helped |
| Add phase I dual-phase structure (ablation) | DGCNN encoder and MLP decoder completion model · Module ablation | Chamfer Distance: 1.850 -> 1.465 | Helped |
| Add the multi-patch generator (ablation) | Dual-phase CDPNet · Module ablation | Chamfer Distance: 1.465 -> 1.103 | Helped |
| Add feature fusion and style-code injection (ablation) | Dual-phase CDPNet with multi-patch generator · Module ablation | Chamfer Distance: 1.103 -> 1.066 | Helped |
| Vary the number of patch generators from 8 to 4 (ablation) | CDPNet · Airplane | Chamfer Distance: 0.482 -> 0.476 | Helped |
| Increase patch generators from 8 to 16 or 32 (ablation) | CDPNet · Airplane | Chamfer Distance: 0.482 -> 0.487; 0.482 -> 0.512 | Hurt |
| Vary the number of patch generators (ablation) | CDPNet · Cabinet, Car, and Watercraft | Chamfer Distance: Cabinet: 1.407, 1.357, 1.381, 1.433; Car: 1.493, 1.470, 1.502, 1.635; Watercraft: 0.543, 0.499, 0.508, 0.571 | Mixed |
Surprising findings
- PoinTr, a strong transformer baseline, performs poorly (CD 2.404), which the authors attribute to overfitting at 2048-point resolution it was not designed for.
Limitations
- Training used only about one third of the ShapeNet-ViPC training data due to computational overhead.
- Performance depends on the view: worst single views give noticeably higher CD than best views.
- Adding more patch generators beyond M=8 degrades results, likely from overfitting.
Experimental setup
Datasets: ShapeNet-ViPC
Benchmarks: ShapeNet-ViPC
Metrics: Chamfer Distance F-Score
Baselines: FoldingNet PCN TopNet MSN GRNet PoinTr XMFNet
Builds on: MSN ViPC DGCNN StyleGAN PCN
Closest papers by meaning
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.90)
- Cross-modal Learning for Image-Guided Point Cloud Shape Completion (NeurIPS 2022 · similarity 0.90)
- DMF-Net: Image-Guided Point Cloud Completion with Dual-Channel Modality Fusion and Shape-Aware Upsampling Transformer (AAAI 2025 · similarity 0.89)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.88)
- CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers (AAAI 2024 · similarity 0.87)
- PointAttN: You Only Need Attention for Point Cloud Completion (AAAI 2024 · 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.