Research map / Continual learning / Domain and prompt-based
A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability
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 a versatile CTTA framework that builds reliable pseudo-labels via self-adaptive thresholds, source-knowledge calibration, and diversity weighting, plus a soft-weighted parameter alignment to limit forgetting, tuning only BN layers.
Problem
Continual test-time domain adaptation must improve performance on the current unlabeled target domain while avoiding catastrophic forgetting and loss of generalization to future domains, which is hurt by noisy pseudo-labels.
Approach
Combines class-wise adaptive confidence thresholds to split pseudo-labels into reliable and unreliable sets, calibrates unreliable labels using nearest neighbours found with source-model features, weights samples by a diversity-and-certainty score from an EMA of predictions, and aligns BN parameters to the source model with depth-dependent soft weights.
Key results
- On CIFAR10-to-CIFAR10C with ResNet, the method reaches 14.5% mean error versus 16.2% for ROID and 16.5% for CoTTA (source 43.5%).
- With ViT-base on CIFAR10-to-CIFAR10C, mean error is 8.9% versus 11.4% for ROID.
- On ImageNet-to-ImageNet-C it reports 51.4% (ResNet) and 41.3% (ViT) average error, improving over competing methods per the text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ROID with the proposed method | ResNet continual test-time adaptation · CIFAR10-to-CIFAR10C, severity 5 | mean error rate (%): 16.2 -> 14.5 | Helped |
| replace ROID with the proposed method | ViT-base continual test-time adaptation · CIFAR10-to-CIFAR10C, severity 5 | mean error rate (%): 11.4 -> 8.9 | Helped |
| replace ROID with the proposed method | ResNet continual test-time adaptation · CIFAR100-to-CIFAR100C, severity 5 | mean error rate (%): 29.3 -> 27.9 | Helped |
| replace ROID with the proposed method | ViT-base continual test-time adaptation · CIFAR100-to-CIFAR100C, severity 5 | mean error rate (%): 25.6 -> 24.0 | Helped |
| replace ROID with the proposed method | ResNet continual test-time adaptation · ImageNet-to-ImageNet-C, severity 5 | average error rate (%): 54.5 -> 51.4 | Helped |
| replace ViDA with the proposed method | ViT-base continual test-time adaptation · ImageNet-to-ImageNet-C, severity 5 | average error rate (%): 43.4 -> 42.7 | Helped |
| replace ROID with the proposed method | ResNet continual test-time adaptation · CIFAR10-to-CIFAR10C, gradually changing | average error rate (%): 6.1 -> 5.7 | Helped |
| add self-adaptive threshold label selection and discard unreliable labels (ablation) | continual test-time adaptation · CIFAR10-to-CIFAR10C | mean error rate (%): 20.4 -> 17.7 | Helped |
| add calibration with source knowledge (ablation) | SST · CIFAR10-to-CIFAR10C | mean error rate (%): 17.7 -> 17.4 | Helped |
| add diversity with prior distribution (ablation) | SST · CIFAR10-to-CIFAR10C | mean error rate (%): 17.7 -> 16.1 | Helped |
| add CSK and DPD (ablation) | SST · CIFAR10-to-CIFAR10C | mean error rate (%): 17.7 -> 15.8 | Helped |
| add CSK and soft-weighted parameter alignment (ablation) | SST · CIFAR10-to-CIFAR10C | mean error rate (%): 17.7 -> 15.6 | Helped |
Surprising findings
- Adapting only normalization layers is reported to be limited on harder datasets such as CIFAR100C and ImageNet-C, where gains are smaller than on CIFAR10C.
Limitations
- Gains are smaller on CIFAR100-C and ImageNet-C, attributed to limited capacity of tuning only normalization layers.
- Relies on a source-model feature similarity for calibrating unreliable pseudo-labels, with hyperparameters (lambda1=0.1, K=3) chosen empirically.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet CIFAR10-C CIFAR100-C ImageNet-C
Benchmarks: CIFAR10-to-CIFAR10C CIFAR100-to-CIFAR100C ImageNet-to-ImageNet-C
Metrics: classification error rate (%)
Baselines: Source BN Stats Adapt Pseudo-Label TENT-continual CoTTA NOTE RoTTA RMT ROID ViDA VDP
Builds on: CoTTA ROID TENT NOTE
Closest papers by meaning
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.90)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.89)
- Continual-MAE: Adaptive Distribution Masked Autoencoders for Continual Test-Time Adaptation (CVPR 2024 · similarity 0.89)
- BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation (ICML 2024 · similarity 0.88)
- Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation (NeurIPS 2025 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 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.