Research map / Point clouds / Shape and surface completion
PointCFormer: A Relation-Based Progressive Feature Extraction 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 PointCFormer, a transformer for point cloud completion that adds a relation-based local geometry perception module, a progressive self-attention/local feature extractor, and a latent point-proxy correction module.
Problem
Point cloud completion loses global structure under sparse down-sampling and local feature extractors (kNN/EdgeConv) capture coarse local geometry.
Approach
Weights each k-NN neighbor's contribution using relation metrics from 3D spatial offsets (R1) and from deviation of edge vectors from the local mean edge (R2), combined with a progressive dense-to-sparse extractor that alternates self-attention and local relation queries, plus a correction module that re-injects dense input information into point proxies.
Key results
- On PCN, PointCFormer reaches average CD-l1 6.41 and F-Score@1% 0.855, versus 6.53 and 0.845 for AdaPoinTr (the strongest baseline in the table).
- On ShapeNet-55, average CD-l2 is 0.73 (best), while F-Score@1% of 0.499 trails AdaPoinTr's 0.503.
- Ablation on PCN: the local relation module alone lowers CD-l1 from 6.92 to 6.47; the full model with all three modules reaches 6.41.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PointCFormer instead of AdaPoinTr | Point cloud completion · PCN | Avg CD-ℓ1: 6.53 -> 6.41 | Helped |
| Use PointCFormer instead of AdaPoinTr | Point cloud completion · PCN | F-Score@1%: 0.845 -> 0.855 | Helped |
| Use PointCFormer instead of PointAttN | Point cloud completion · ShapeNet-55 | CD-ℓ2-Avg: 0.77 -> 0.73 | Helped |
| Use PointCFormer instead of AdaPoinTr | Point cloud completion · ShapeNet-55 | F-Score@1%: 0.503 -> 0.499 | Hurt |
| Use PointCFormer instead of AdaPoinTr | Point cloud completion · ShapeNet-34 seen categories | CD-ℓ2-Avg: 0.73 -> 0.66 | Helped |
| Use PointCFormer instead of AdaPoinTr | Point cloud completion · ShapeNet-34 unseen categories | CD-ℓ2-Avg: 1.23 -> 1.12 | Helped |
| Add LGRP (ablation) | PointCFormer feature extractor · PCN | CD-ℓ1: 6.92 -> 6.47 | Helped |
| Add LGRP (ablation) | PointCFormer feature extractor · PCN | F-Score@1%: 0.810 -> 0.840 | Helped |
| Add GSP only (ablation) | PointCFormer feature extractor · PCN | CD-ℓ1: 6.92 -> 6.85 | Helped |
| Add GSP only (ablation) | PointCFormer feature extractor · PCN | F-Score@1%: 0.810 -> 0.822 | Helped |
| Add CM only (ablation) | PointCFormer · PCN | CD-ℓ1: 6.92 -> 6.87 | Helped |
| Add CM only (ablation) | PointCFormer · PCN | F-Score@1%: 0.810 -> 0.819 | Helped |
Limitations
- F-Score on ShapeNet-55 and ShapeNet-34 is slightly below AdaPoinTr.
- Efficiency claims are not quantified with runtime or parameter counts in the text.
Experimental setup
Datasets: PCN ShapeNet-55 ShapeNet-34 Unseen-21 KITTI ShapeNetCars
Benchmarks: PCN ShapeNet-55 ShapeNet-34/Unseen-21 KITTI
Metrics: CD-l1 CD-l2 F-Score@1% fidelity distance minimum matching distance
Baselines: FoldingNet PCN GRNet PoinTr SnowflakeNet SeedFormer AnchorFormer PointAttN AdaPoinTr 3DMambaComplete
Builds on: PoinTr AdaPoinTr DGCNN PointAttN SeedFormer
Closest papers by meaning
- AnchorFormer: Point Cloud Completion From Discriminative Nodes (CVPR 2023 · similarity 0.90)
- PointAttN: You Only Need Attention for Point Cloud Completion (AAAI 2024 · similarity 0.89)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.89)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.88)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.88)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 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.