Research map / Point clouds / Shape and surface completion
CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers
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 CRA-PCN, a coarse-to-fine point cloud completion network whose up-sampling blocks combine recursive inter-level and intra-level Cross-Resolution Transformers based on local vector attention.
Problem
Point cloud completion: recovering complete point clouds from incomplete partial scans, where existing coarse-to-fine methods use only inter-level cross-resolution aggregation and neglect intra-level aggregation.
Approach
A unified Cross-Resolution Transformer performs recursive multi-scale local-attention aggregation between a query and a support point cloud; switching inputs yields inter-level (support = earlier generated cloud) or intra-level (support = same cloud) aggregation, and both are combined in each up-sampling block.
Key results
- On PCN, CRA-PCN achieves average CD-L1 of 6.39 x10^3, improving on the second-best AnchorFormer (6.59) by 0.20.
- On MVP, CRA-PCN reaches CD-L2 5.33 x10^4 and F-Score 0.529, versus VRCNet's 5.96 and 0.499.
- On ShapeNet-55, CRA-PCN reports average CD-L2 of 0.85 x10^3, beating SeedFormer (0.92).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CRA-PCN instead of AnchorFormer | point cloud completion · PCN | L1 Chamfer Distance × 10^3: 6.59 -> 6.39 | Helped |
| use CRA-PCN instead of SeedFormer | point cloud completion · ShapeNet-55 | CD-Avg: 0.92 -> 0.85 | Helped |
| use CRA-PCN instead of SeedFormer | point cloud completion · ShapeNet-34 seen categories | CD-Avg: 0.83 -> 0.76 | Helped |
| use CRA-PCN instead of SeedFormer | point cloud completion · ShapeNet-34 unseen categories | CD-Avg: 1.34 -> 1.24 | Helped |
| use CRA-PCN instead of VRCNet | point cloud completion · MVP | L2 Chamfer Distance × 10^4: 5.96 -> 5.33 | Helped |
| use CRA-PCN instead of VRCNet | point cloud completion · MVP | F-Score: 0.499 -> 0.529 | Helped |
| add inter-level and intra-level CRT (ablation) | CRA-PCN up-sampling block · PCN | CD-ℓ1: 7.85 -> 6.39 | Helped |
| add inter-level CRT (ablation) | CRA-PCN up-sampling block with intra-level CRT · PCN | CD-ℓ1: 6.92 -> 6.39 | Helped |
| add intra-level CRT (ablation) | CRA-PCN up-sampling block with inter-level CRT · PCN | CD-ℓ1: 6.80 -> 6.39 | Helped |
| use one inter-level and one intra-level CRT instead of two intra-level CRTs (ablation) | CRA-PCN up-sampling block · PCN | CD-ℓ1: 6.74 -> 6.39 | Helped |
| use one inter-level and one intra-level CRT instead of two inter-level CRTs (ablation) | CRA-PCN up-sampling block · PCN | CD-ℓ1: 6.61 -> 6.39 | Helped |
| use the standard inter-level then intra-level CRT order instead of reversing it (ablation) | CRA-PCN up-sampling block · PCN | CD-ℓ1: 6.54 -> 6.39 | Helped |
Limitations
- Latency (0.469 s) is higher than SeedFormer (0.297 s) and memory usage (4459 MB) is higher than FBNet and AnchorFormer, so efficiency is not uniformly better.
- Gains over strong baselines are modest on some benchmarks (e.g. 0.20 CD on PCN versus AnchorFormer).
Experimental setup
Datasets: PCN ShapeNet-55 ShapeNet-34 MVP
Benchmarks: PCN ShapeNet-55 ShapeNet-34 MVP
Metrics: L1 Chamfer Distance L2 Chamfer Distance F-Score latency memory
Baselines: FoldingNet TopNet PCN GRNet PoinTr SnowflakeNet FBNet ProxyFormer SeedFormer AnchorFormer PFNet MSN
Builds on: PoinTr SeedFormer Point Transformer PointNet++ FoldingNet
Closest papers by meaning
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.88)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.88)
- CDPNet: Cross-Modal Dual Phases Network for Point Cloud Completion (AAAI 2024 · similarity 0.87)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.86)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 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.