Research map / Label-efficient and robust learning / Test-time adaptation
Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online 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 DPLOT, an online test-time adaptation framework that entropy-minimizes only selected domain-specific blocks and trains all parameters with EMA-teacher pseudo-labels averaged over horizontally flipped image pairs.
Problem
Online test-time adaptation to corrupted data suffers from error accumulation because pseudo-labels from the teacher degrade over long non-stationary test sequences; existing methods update BN layers and use strong augmentations that create domain gaps.
Approach
Select domain-specific blocks before deployment by measuring prototype similarity change under entropy minimization on Gaussian-noised source data, entropy-minimize only those blocks at test time, and generate pseudo-labels by averaging an EMA teacher's predictions on a test image and its horizontal flip instead of strong augmentations.
Key results
- In the gradual CIFAR10-C setting with WRN40-2A, DPLOT reaches 10.2% mean error over severity levels 1-5 vs 10.9% for RMT.
- In the continual CIFAR100-C setting with ResNext-29A, DPLOT achieves 27.8% average error vs 29.0% for RMT and 32.5% for CoTTA.
- Ablation on gradual CIFAR10-C (WRN28-10) shows removing paired-view consistency raises level-5 error from 8.8% to 11.3%, and replacing block selection with BN updates raises it to 9.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPLOT instead of the strongest listed baseline | CIFAR100-C continual, ResNext-29A · CIFAR100-C continual | error rate: 29.0 -> 27.8 | Helped |
| Use DPLOT instead of the strongest listed baseline | CIFAR100-C continual, WRN40-2A · CIFAR100-C continual | error rate: 34.3 -> 31.8 | Helped |
| Use DPLOT instead of the strongest listed baseline | ImageNet-C continual, ResNet-50 · ImageNet-C continual | error rate: 58.8 -> 60.2 | Hurt |
| Use DPLOT instead of the strongest listed baseline | ImageNet-C continual, ResNet-50A · ImageNet-C continual | error rate: 46.3 -> 44.6 | Helped |
| Use DPLOT rather than RMT | CIFAR10-C gradual, WRN28-10 · CIFAR10-C gradual | error rate: 9.3 -> 8.8 | Helped |
| Use DPLOT rather than RMT or SAR | CIFAR10-C gradual, WRN40-2A and ResNet-18A · CIFAR10-C gradual | error rate: 7.7 -> 7.2; 7.6 -> 6.3 | Helped |
| Use DPLOT rather than RMT or SAR | CIFAR100-C gradual, ResNext-29A and WRN40-2A · CIFAR100-C gradual | error rate: 26.4 -> 23.9; 30.7 -> 27.3 | Helped |
| Use DPLOT rather than CoTTA | ImageNet-C gradual, ResNet-50 and ResNet-50A · ImageNet-C gradual | error rate: 38.8 -> 37.2; 32.0 -> 30.8 | Helped |
| Add paired-view consistency (ablation) | DPLOT · CIFAR10-C, CIFAR100-C, and ImageNet-C gradual (level 5) | error rate: 11.3 -> 8.8; 25.9 -> 23.9; 37.4 -> 30.8 | Helped |
| Add the EMA teacher (ablation) | DPLOT · CIFAR10-C, CIFAR100-C, and ImageNet-C gradual (level 5) | error rate: 12.5 -> 8.8; 25.8 -> 23.9; 37.9 -> 30.8 | Helped |
| Add domain-specific block selection (ablation) | DPLOT · CIFAR10-C, CIFAR100-C, and ImageNet-C gradual (level 5) | error rate: 19.3 -> 8.8; 66.9 -> 23.9; 38.5 -> 30.8 | Helped |
Surprising findings
- On ResNet-50 ImageNet-C in the continual setting, DPLOT (60.2%) is worse than EATA (58.8%), RMT (59.8%), and SAR (61.9%)-style baselines, so it does not win across all architectures.
- Selecting blocks for entropy minimization is damaging when the threshold is too low (gamma=0.0 gives 80.7% error on CIFAR10-C), showing that updating domain-invariant blocks causes model collapse.
Limitations
- Does not win on every architecture or benchmark; underperforms on ResNet-50 for ImageNet-C in the continual setting.
- Block selection relies on a threshold gamma that must be chosen; very low values cause collapse, and the method uses Gaussian noise as a proxy for domain shift during selection.
Experimental setup
Datasets: CIFAR10-C CIFAR100-C ImageNet-C CIFAR10 CIFAR100 ImageNet
Benchmarks: CIFAR10-C CIFAR100-C ImageNet-C
Metrics: classification error rate (%)
Baselines: BN-1 TENT AdaContrast CoTTA EATA SAR RMT
Builds on: TENT CoTTA RMT EATA SAR
Closest papers by meaning
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.89)
- Test-time Correlation Alignment (ICML 2025 · similarity 0.88)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Multi-Label Test-Time Adaptation with Bound Entropy Minimization (ICLR 2025 · similarity 0.87)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.87)
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.