Research map / Point clouds / Shape and surface completion
PointMAC: Meta-Learned Adaptation for Robust Test-Time 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 PointMAC, a MAML-based meta-auxiliary test-time adaptation framework for point cloud completion that adapts a shared encoder per sample using masked-reconstruction and denoising auxiliary losses.
Problem
Point cloud completion models perform static inference and rely on training-set inductive biases, so they generalize poorly to novel structures and sensor-induced distortions at test time.
Approach
Treats each point cloud as its own domain: during meta-training, MAML simulates inner-loop auxiliary adaptation (stochastic masked reconstruction and artifact denoising) with outer-loop primary completion supervision, plus an Adaptive λ-Calibration that learns auxiliary loss weights; at inference only the shared encoder is adapted with a few gradient steps on the auxiliary losses while the decoder
Key results
- On PCN, PointMAC reports CD-l1 avg 6.33 (x1000), best among compared methods (CRA-PCN 6.39).
- On MVP, CD-l2 5.24 (x10^4) and F1 0.537, versus CRA-PCN 5.33 and 0.529.
- On KITTI (trained on ShapeNetCars), Fidelity drops from 0.151 to 0.135 and MMD from 0.508 to 0.477 relative to the non-adapted model/prior baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Bi-Aux Units (ablation) | PointMAC baseline · PCN | CD-ℓ1: 6.62 -> 6.42 | Helped |
| add Bi-Aux Units (ablation) | PointMAC baseline · ShapeNet-55 | CD-ℓ2: 0.90 -> 0.85 | Helped |
| remove Token Synergy Integrator (ablation) | PointMAC full model · PCN | CD-ℓ1: 6.33 -> 6.49 | Hurt |
| remove Token Synergy Integrator (ablation) | PointMAC full model · ShapeNet-55 | CD-ℓ2: 0.83 -> 0.86 | Hurt |
| remove Adaptive λ-Calibration (ablation) | PointMAC full model · PCN | CD-ℓ1: 6.33 -> 6.38 | Hurt |
| remove Adaptive λ-Calibration (ablation) | PointMAC full model · ShapeNet-55 | CD-ℓ2: 0.83 -> 0.84 | Hurt |
| add TTA with K = 1 update (ablation) | Bi-Aux Units model · PCN | CD-ℓ1: 6.42 -> 6.40 | Helped |
| add TTA with K = 1 update (ablation) | Bi-Aux Units model · ShapeNet-55 | CD-ℓ2: 0.85 -> 0.84 | Helped |
| increase TTA updates from K = 3 to K = 5 (ablation) | Bi-Aux Units + TTA · PCN | CD-ℓ1: 6.33 -> 6.28 | Helped |
| increase TTA updates from K = 3 to K = 5 (ablation) | Bi-Aux Units + TTA · ShapeNet-55 | CD-ℓ2: 0.83 -> 0.81 | Helped |
| use PointMAC instead of CRA-PCN | Point cloud completion · MVP | CD-ℓ2 (×10^4): 5.33 -> 5.24 | Helped |
| use PointMAC instead of CRA-PCN | Point cloud completion · MVP | F1: 0.529 -> 0.537 | Helped |
Limitations
- Experiments are limited to single-object completion; scene-level extension is untested.
- Test-time adaptation adds per-sample gradient steps, increasing inference cost.
- Real-world KITTI evaluation lacks paired ground truth and relies on Fidelity/MMD metrics.
Experimental setup
Datasets: PCN ShapeNet-55 ShapeNet-34 MVP ShapeNetCars
Benchmarks: PCN ShapeNet-55 ShapeNet-34 MVP KITTI
Metrics: CD-l1 CD-l2 F-Score@1% Fidelity MMD
Baselines: FoldingNet PCN SnowflakeNet PoinTr SeedFormer ProxyFormer EINet CRA-PCN
Builds on: MAML Dis-PU PoinTr SeedFormer
Closest papers by meaning
- SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.88)
- ASHF-Net: Adaptive Sampling and Hierarchical Folding Network for Robust Point Cloud Completion (AAAI 2021 · similarity 0.88)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.86)
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.86)
- Point Cloud Matters: Rethinking the Impact of Different Observation Spaces on Robot Learning (NeurIPS 2024 · similarity 0.86)
- PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 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.