Research map / Point clouds / Shape and surface completion
PointAttN: You Only Need Attention 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 PointAttN, a kNN-free encoder-decoder for point cloud completion built only from MLPs, FPS, and cross/self-attention units (GDP and SFA), reaching state-of-the-art on several benchmarks.
Problem
Point cloud completion from partial scans; existing methods rely on kNN/ball-query local partitions that are sensitive to point density and have limited receptive fields.
Approach
Replace explicit kNN local partitioning with implicit local aggregation via cross-attention (GDP, between a point set and its FPS-downsampled version) and self-attention (SFA, for global context), arranged in a coarse-to-fine encoder-decoder with only MLP, FPS and attention layers.
Key results
- On Completion3D, average CD-L2 of 6.63 vs. 7.60 for SnowflakeNet (about 12.8% reduction).
- On PCN, average CD-L1 of 6.84, second only to SeedFormer (6.74); on ShapeNet-55 it reduces CD-L2 versus PoinTr and SeedFormer at simple, moderate and hard levels.
- On KITTI, MMD of 0.504 (best) with Fidelity 0.672, versus SeedFormer MMD 0.516 and Fidelity 0.151.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GDP units to the feature extractor (ablation) | PointAttN · Completion3D | CD-L2 (Avg): 7.42 -> 6.63 | Helped |
| add SFA units to the feature extractor (ablation) | PointAttN · Completion3D | CD-L2 (Avg): 6.81 -> 6.63 | Helped |
| use the PointAttN feature extractor with the FoldingNet backbone (ablation) | FoldingNet · Completion3D | CD-L2 (Avg): 19.07 -> 8.28 | Helped |
| replace the point generator with the SnowflakeNet SPD decoder (ablation) | PointAttN · Completion3D | CD-L2 (Avg): 6.63 -> 7.31 | Hurt |
| replace the feature extractor with PointNet++ (ablation) | PointAttN · Completion3D | CD-L2 (Avg): 6.63 -> 7.23 | Hurt |
| use PointAttN instead of SeedFormer | point cloud completion · PCN | CD-L1 (Avg): 6.74 -> 6.84 | Hurt |
| use PointAttN instead of SeedFormer | point cloud completion · ShapeNet-34 21 unseen categories, hard | CD-L2 (CD-H): 2.35 -> 2.23 | Helped |
| use PointAttN instead of SeedFormer | point cloud completion · KITTI | MMD: 0.516 -> 0.504 | Helped |
Surprising findings
- On PCN, PointAttN is slightly worse than SeedFormer on average CD (6.84 vs. 6.74), despite being better on several categories.
Limitations
- Slightly behind SeedFormer on PCN average CD.
- Requires a larger parameter count and inference time than some lightweight baselines (e.g., SnowflakeNet's 13.45 ms).
Experimental setup
Datasets: Completion3D PCN ShapeNet-55 ShapeNet-34 KITTI ShapeNetCars
Benchmarks: Completion3D PCN ShapeNet-55/34 KITTI
Metrics: Chamfer distance (CD-L1/L2) MMD Fidelity parameters inference time VRAM
Baselines: GRNet PoinTr SeedFormer SnowflakeNet FBNet PMP-Net++ VRC-Net SA-Net FoldingNet GTNet SCRN SoftPool++
Builds on: Transformer PointNet++ SnowflakeNet PoinTr
Closest papers by meaning
- Point Cloud Completion by Skip-Attention Network With Hierarchical Folding (CVPR 2020 · similarity 0.90)
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.89)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.89)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.88)
- Orthogonal Dictionary Guided Shape Completion Network for Point Cloud (AAAI 2024 · similarity 0.87)
- CDPNet: Cross-Modal Dual Phases Network 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.