Research map / Continual learning / Domain and prompt-based
ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation
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 ViDA, a visual domain adapter with paired high-rank and low-rank branches plus an uncertainty-driven Homeostatic Knowledge Allotment, for continual test-time adaptation of pre-trained vision models.
Problem
Continual test-time adaptation (CTTA) must adapt a source-pretrained model to continually changing target domains while avoiding error accumulation from noisy pseudo labels and catastrophic forgetting.
Approach
Inject parallel high-rank (domain-specific) and low-rank (domain-shared) adapters into pretrained layers, and dynamically weight them per sample using MC-dropout uncertainty (HKA); adapters are re-parameterized into the model at inference so no extra parameters remain.
Key results
- On ImageNet-to-ImageNet-C with ViT-base, ViDA reaches 43.4% mean error vs 55.8% source and 49.9% (VDP) best prior, a 12.4% gain over source.
- On Cityscapes-to-ACDC (Segformer-B5), ViDA achieves 61.9 mean mIoU over three rounds, +5.2 over source, and 3.3 above the prior SOTA CTTA method.
- Ablation on ImageNet-C: high-rank ViDA alone reduces error 55.8→50.7, low-rank alone →51.2, both with HKA →43.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ViDA instead of the strongest listed baseline VDP | ViT-base · ImageNet-to-ImageNet-C | classification error rate (%): 50.0 -> 43.4 | Helped |
| Use ViDA instead of TENT | ViT-base · CIFAR10-to-CIFAR10C | mean classification error rate (%): 23.5 -> 20.7 | Helped |
| Use ViDA instead of TENT | ViT-base · CIFAR100-to-CIFAR100C | mean classification error rate (%): 32.1 -> 27.3 | Helped |
| Use ViDA instead of CoTTA | SegFormer-B5 · Cityscapes-to-ACDC | mean mIoU: 58.6 -> 61.9 | Helped |
| Add high-rank ViDA (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 55.8 -> 50.7 | Helped |
| Add low-rank ViDA (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 55.8 -> 51.2 | Helped |
| Combine high- and low-rank ViDAs (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 55.8 -> 45.6 | Helped |
| Add HKA to the high-/low-rank ViDA pair (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 45.6 -> 43.4 | Helped |
| Invert HKA scale factors (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | classification error rate (%): 43.4 -> 46.3 | Hurt |
| Replace the high-/low-rank pair with two low-rank adapters (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 45.6 -> 47.7 | Hurt |
| Replace the high-/low-rank pair with two high-rank adapters (ablation) | ViDA on ViT-base · ImageNet-to-ImageNet-C | mean classification error rate (%): 45.6 -> 48.9 | Hurt |
| Randomly initialize adapter parameters rather than pre-train them on the source (ablation) | ViDA with SegFormer-B5 · Cityscapes-to-ACDC | mean mIoU: 61.1 -> 60.5 | Hurt |
Surprising findings
- Inverting the HKA uncertainty weighting increases error (46.3 vs 43.4), showing the direction of allotment matters.
- Using SAM pretrained encoder with a classification head degraded classification CTTA performance relative to the source model in the main ViT setting, which the authors called surprising.
Limitations
- Relies on a teacher-student EMA framework and MC-dropout uncertainty with a hand-set threshold (0.2).
- Gains on some corruption types are not consistent; some results are worse than prior methods on individual domains.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C Cityscapes ACDC
Benchmarks: CIFAR10-to-CIFAR10C CIFAR100-to-CIFAR100C ImageNet-to-ImageNet-C Cityscapes-to-ACDC
Metrics: error rate (%) mIoU gain over source
Baselines: Source Tent CoTTA EcoTTA VDP DePT RMT SATA
Builds on: CoTTA Tent AdaptFormer EcoTTA
Closest papers by meaning
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.90)
- A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability (CVPR 2024 · similarity 0.89)
- PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation (AAAI 2025 · similarity 0.88)
- 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.