Research map / Point clouds / Shape and surface completion
DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization
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 DC-PCN, a transformer encoder-decoder point cloud completion network with encoder and decoder vector-quantized codebooks plus a quantized information exchange mechanism, reporting state-of-the-art CD and F-Score on PCN, ShapeNet_Part and ShapeNet34.
Problem
Point cloud completion is ambiguous because point clouds sampled from the same 3D surface vary, which yields inconsistent latent representations and sub-optimal completions.
Approach
A dual-codebook design quantizes shallow (encoder-side) and deep (decoder-side) point features with separate VQ codebooks, and a quantized information exchange (code deduplication, distribution re-targeting via MLP, and similarity-weighted code merging) links the two codebooks so features from the same surface map to consistent codes.
Key results
- On PCN, average CD-l1 of 6.46 (x10^3) and F-Score@1 of 0.850, best among compared methods, improving on AdaPoinTr's 6.53 average CD-l1 by about 0.07.
- On ShapeNet_Part, average CD-l2 of 5.59 (x10^4) versus AdaPoinTr 6.10 and PoinTr 6.26.
- Ablation on PCN: encoder-codebook lowers CD-l1 from 6.53 to 6.48; full dual-codebook with QIE reaches 6.46, while a shared codebook raises it to 6.685.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add encoder-codebook (shallow-level quantization) (ablation) | DC-PCN · PCN | CD-l1 (x10^3): -0.05 | Helped |
| add decoder-codebook (deep-level quantization) (ablation) | DC-PCN · PCN | CD-l1 (x10^3): -0.01 | Helped |
| add quantized information exchanging (QIE) mechanism to dual codebook (ablation) | DC-PCN · PCN | CD-l1 (x10^3): -0.01 | Helped |
| replace dual codebook with a single shared codebook (ablation) | DC-PCN · PCN | CD-l1 (x10^3): +0.15 vs. no codebook | Hurt |
| full dual-codebook design (EC+DC) vs. no codebook (ablation) | DC-PCN · PCN | CD-l1 (x10^3): -0.06 | Helped |
| DC-PCN vs. previous best (AdaPoinTr) on PCN | DC-PCN · PCN | average CD-l1 (x10^3): -0.07 | Helped |
| DC-PCN vs. AdaPoinTr on ShapeNet_Part | DC-PCN · ShapeNet_Part | average CD-l2 (x10^4): 5.6 vs 6.3 (AdaPoinTr) | Helped |
Surprising findings
- A single shared codebook for shallow and deep features degrades CD-l1 to 6.685, worse than no codebook at all (6.53).
- The decoder-codebook alone gives only a marginal gain (6.52 vs 6.53 CD-l1).
Limitations
- Performance depends on codebook hyperparameters such as codebook size, which must be tuned per dataset.
- Codebook sizes were chosen as optimal values from parameter sweeps, with no adaptive mechanism.
Experimental setup
Datasets: PCN ShapeNet_Part ShapeNet34 KITTI
Benchmarks: PCN ShapeNet_Part ShapeNet34 KITTI
Metrics: CD-l1 CD-l2 F-Score@1 MMD
Baselines: FoldingNet AtlasNet PCN TopNet MSN GRNet CRN NSFA SnowFlake PoinTr AdaPoinTr SeedFormer
Builds on: VQ-VAE ShapeFormer AutoSDF PoinTr FoldingNet
Closest papers by meaning
- Point Cloud Completion by Skip-Attention Network With Hierarchical Folding (CVPR 2020 · similarity 0.90)
- CDPNet: Cross-Modal Dual Phases Network for Point Cloud Completion (AAAI 2024 · similarity 0.90)
- PointAttN: You Only Need Attention for Point Cloud Completion (AAAI 2024 · similarity 0.89)
- Orthogonal Dictionary Guided Shape Completion Network for Point Cloud (AAAI 2024 · similarity 0.88)
- CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers (AAAI 2024 · similarity 0.88)
- Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.88)
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.