Research map / Continual learning / Domain and prompt-based
PCoTTA: Continual Test-Time Adaptation for Multi-Task Point Cloud Understanding
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 PCoTTA, a continual test-time adaptation framework for multi-task point cloud understanding (reconstruction, denoising, registration) using a prototype bank with source and learnable prototypes, Gaussian-splatted graph attention feature shifting, and contrastive prototype repulsion.
Problem
Test-time adaptation of point cloud models to continually changing target domains fails due to catastrophic forgetting and error accumulation, and existing CoTTA methods are task-specific or designed for 2D images.
Approach
Mixes source prototypes with learnable target prototypes via a similarity-balancing factor, shifts test features using Gaussian-splatted graph attention over prototype-pair nodes, and uses a contrastive repulsion loss to keep learnable prototypes distinct, all within one unified masked-point-modeling model without pseudo-labels or teacher-student training.
Key results
- PCoTTA reaches CD of 6.3 (reconstruction), 21.4 (denoising), 15.4 (registration) on ModelNet40 in round 1, versus 31.2 / 44.0 / 34.3 for the best CoTTA-style baseline RMT.
- Ablation on ModelNet40 shows reconstruction CD falling from 78.9 (baseline) to 23.5 (APM only) to 14.6 (+GSFS) to 5.4 (+CPR).
- Reports 0.06 s runtime and 28.91M parameters, the lowest parameter count among compared CoTTA methods in its efficiency table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ModelNet40 round 3 reconstruction | CD (× 10⁻³): 30.1 -> 5.4 | Helped |
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ModelNet40 round 3 denoising | CD (× 10⁻³): 41.6 -> 18.6 | Helped |
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ModelNet40 round 3 registration | CD (× 10⁻³): 32.8 -> 12.1 | Helped |
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ScanObjectNN round 3 reconstruction | CD (× 10⁻³): 40.6 -> 8.2 | Helped |
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ScanObjectNN round 3 denoising | CD (× 10⁻³): 51.1 -> 25.2 | Helped |
| Use PCoTTA instead of the strongest reported CoTTA baseline | Continual multi-task point cloud adaptation · ScanObjectNN round 3 registration | CD (× 10⁻³): 34.7 -> 19.3 | Helped |
| Add APM to the baseline (ablation) | PCoTTA · ModelNet40 reconstruction | CD (× 10⁻³): 78.9 -> 23.5 | Helped |
| Add APM to the baseline (ablation) | PCoTTA · ModelNet40 denoising | CD (× 10⁻³): 123.6 -> 37.4 | Helped |
| Add APM to the baseline (ablation) | PCoTTA · ModelNet40 registration | CD (× 10⁻³): 110.6 -> 30.1 | Helped |
| Add GSFS to APM (ablation) | PCoTTA · ModelNet40 reconstruction | CD (× 10⁻³): 23.5 -> 14.6 | Helped |
| Add CPR to APM and GSFS (ablation) | PCoTTA · ModelNet40 reconstruction | CD (× 10⁻³): 14.6 -> 5.4 | Helped |
| Set learnable prototype quantity to 0 instead of 2 (ablation) | PCoTTA · ModelNet40 reconstruction | CD (× 10⁻³): 5.4 -> 15.5 | Hurt |
Surprising findings
- Reported efficiency table lists PCoTTA at 12.11 GFLOPs versus CoTTA's 24.26, while the paper frames the method as the most efficient; the runtime comparison is only against CoTTA and ViDA.
Limitations
- Evaluation is limited to two target domains (ModelNet40 and ScanObjectNN) and three tasks on a single benchmark constructed by the authors.
- Relies on a masked point modeling backbone pre-trained on the source domains; gains depend on this pre-training setup.
Experimental setup
Datasets: ModelNet40 ShapeNet ScanNet ScanObjectNN
Benchmarks: ModelNet40 ScanObjectNN
Metrics: Chamfer Distance (CD) runtime (s) FLOPs (G) parameters (M)
Baselines: PointNet DGCNN PCT Pointmixup PointCutMix Baseline PIC AdaBN TENT CoTTA ViDA RMT
Builds on: PIC CoTTA Point-MAE
Closest papers by meaning
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.86)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.85)
- BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation (ICML 2024 · similarity 0.85)
- Hierarchical Knowledge Prompt Tuning for Multi-task Test-Time Adaptation (CVPR 2025 · similarity 0.85)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.85)
- PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion (NeurIPS 2025 · similarity 0.84)
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.