Research map / Point clouds / Shape and surface completion
FSC: Few-point Shape 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 Few-point Shape Completion (FSC), a dual-branch encoder with a two-stage WGAN-based revision network for completing 3D point clouds from as few as 64 input points.
Problem
Point cloud shape completion degrades severely when only a few points (tens of points) are observed, since local detail and global shape information are lost.
Approach
A dual-branch feature extractor (an extensive MLP/PCN-style branch that uses all points plus a transformer-based saliency branch that weights points dynamically), followed by a two-stage WGAN revision of both the latent feature and the coarse point cloud, decoded with PointNet++/external attention and FoldingNet.
Key results
- On 64-point ShapeNet completion, FSC reaches average CD-l1 of 7.63 vs 10.02 for SVDFormer (best prior in Table 1 average).
- On KITTI, FSC achieves MMD 0.239 vs 0.392 for AdaPoinTr (best prior).
- Entropy analysis finds a 64-point input retains 45.51% of the ground-truth FPFH shape information.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FSC instead of the strongest listed baseline | Shape completion · ShapeNet seen categories, 64 points | CD-ℓ1: 11.22 -> 7.89 | Helped |
| Use FSC instead of the strongest listed baseline | Shape completion · ShapeNet seen categories, 2048 points | CD-ℓ1: 7.21 -> 7.02 | Helped |
| Use FSC instead of the strongest listed baseline | Shape completion · Completion3D | CD-ℓ2: 6.68 -> 6.66 | Helped |
| Use FSC instead of the strongest listed baseline | Shape completion · KITTI | MMD: 0.392 -> 0.239 | Helped |
| Use FSC instead of the strongest listed baseline | Shape completion · ShapeNet unseen categories, 64 points | CD-ℓ1: 13.60 -> 12.90 | Helped |
| Add a salient branch to the extensive branch (ablation) | FSC feature extractor · ShapeNet seen categories, 64 points | CD-ℓ1: 10.53 -> 9.78 | Helped |
| Use the full FSC design rather than the extensive-branch configuration (ablation) | FSC · ShapeNet seen categories, 64 points | CD-ℓ1: 10.53 -> 7.89 | Helped |
Surprising findings
- Even 64 randomly sampled points retain about 45% of the shape information measured by FPFH entropy.
Limitations
- Performance drops sharply when input has fewer than 64 points.
- Results still contain deformation and noise on 64-point inputs.
- Unseen-category results are mixed on some classes (e.g., Bookshelf worse than SVDFormer).
Experimental setup
Datasets: ShapeNet KITTI Completion3D
Benchmarks: ShapeNet KITTI Completion3D
Metrics: CD-l1 CD-l2 MMD FPFH Shannon entropy
Baselines: GRNet PCN PoinTr SeedFormer SVDFormer PMP-Net SnowFlake AdaPoinTr
Builds on: PCN FoldingNet PointNet++ WGAN Chamfer distance
Closest papers by meaning
- Learning 3D Shape Completion From Laser Scan Data With Weak Supervision (CVPR 2018 · similarity 0.89)
- Orthogonal Dictionary Guided Shape Completion Network for Point Cloud (AAAI 2024 · similarity 0.87)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.87)
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.87)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.87)
- PointAttN: You Only Need Attention for Point Cloud Completion (AAAI 2024 · 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.