Research map / Label-efficient and robust learning / Test-time adaptation
Improved Self-Training for 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 Improved Self-Training (IST) for fully test-time adaptation, combining crop/flip test-time augmentation, graph-based pseudo-label correction in latent space, and parameter moving average, evaluated on classification, detection and segmentation.
Problem
Self-trained test-time adaptation suffers from unreliable pseudo-labels (poor calibration, augmentation harming fine-grained classes) and unstable model optimization with catastrophic forgetting, under fully test-time adaptation with no source data.
Approach
Uses a simple crop/scale/flip augmentation with NMS-style max-confidence selection instead of averaging, corrects pseudo-labels via label propagation over a K-NN graph of target latent features, and stabilizes updates with an exponential moving average of parameters toward the source model, applied to CLIP as source model.
Key results
- On CIFAR-10C (level 5, WRN-40-2) the online method reaches 10.2% mean error versus 11.5% for TENT and 10.9% for MEMO.
- With CLIP ViT-B/32 on CIFAR10, the offline variant reaches 95.7% top-1 versus 88.8% zero-shot.
- On KITTI fog, SimpAug alone improves mAP@50 from 22.8 to 27.3 and online adaptation reaches 43.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Adapt CLIP with offline IST rather than zero-shot inference | CLIP ViT-B/32 · CIFAR10 | top-1 accuracy (%): 88.8±0.0 -> 95.7±0.0 | Helped |
| Adapt CLIP with offline IST rather than zero-shot inference | CLIP ViT-B/32 · CIFAR100 | top-1 accuracy (%): 61.1±0.0 -> 75.9±0.1 | Helped |
| Adapt CLIP with offline IST rather than zero-shot inference | CLIP ViT-B/32 · Food101 | top-1 accuracy (%): 80.7±0.0 -> 87.1±0.1 | Helped |
| Adapt CLIP with offline IST rather than zero-shot inference | CLIP ViT-B/32 · Cars | top-1 accuracy (%): 58.6±0.0 -> 69.3±0.0 | Helped |
| Adapt CLIP with offline IST rather than zero-shot inference | CLIP ViT-B/32 · ImageNet | top-1 accuracy (%): 59.8±0.0 -> 66.2±0.0 | Helped |
| Use online IST instead of the strongest listed baseline | WRN-40-2 · CIFAR-10C level 5 | top-1 error (%): 9.6±0.00 -> 10.2±0.04 | Mixed |
| Use online IST instead of the strongest listed baseline | ResNet-18 · ImageNet-C level 5 | top-1 error (%): 62.5±0.00 -> 63.4±0.03 | Mixed |
| Use online IST instead of CoTTA | ResNet-18 · ImageNet-C continuous adaptation | top-1 error (%): 67.8±0.3 -> 65.4±0.1 | Helped |
| Use online IST instead of ActMAD | YOLOv3 · KITTI rain | mAP@50: 81.4±0.2 -> 81.6±0.2 | Helped |
| Use online IST instead of ActMAD | YOLOv3 · KITTI fog | mAP@50: 47.7±1.0 -> 43.2±0.3 | Hurt |
| Use online IST instead of BN Adapt | DeepLab-V2 ResNet-101 · CarlaTTA clear2rain | mIoU (%): 71.4 -> 73.9 | Helped |
| Use online IST instead of BN Adapt | DeepLab-V2 ResNet-101 · CarlaTTA dynamic | mIoU (%): 52.6 -> 51.5 | Hurt |
Surprising findings
- Adam and momentum-based optimizers can fluctuate or decline under error accumulation, while vanilla SGD with small learning rate was more stable.
- Complex augmentations (AugMix-based MEMO) gave worse results on class-dense and fine-grained benchmarks than the simple crop/flip strategy.
Limitations
- Large numbers of augmentations require more computation and storage.
- Online performance is below offline adaptation on several CLIP benchmarks because online training uses pseudo-labels less effectively.
Experimental setup
Datasets: CIFAR10 CIFAR100 Food101 Stanford Cars ImageNet CIFAR-10C ImageNet-C KITTI CarlaTTA ImageNet-V2 ImageNet-Sketch ImageNet-A
Benchmarks: CIFAR10 CIFAR100 Food101 Cars ImageNet CIFAR-10C ImageNet-C KITTI CarlaTTA ImageNet-V2 ImageNet-Sketch ImageNet-A
Metrics: top-1 accuracy top-1 error mAP@50 mIoU
Baselines: BN Adapt TENT MEMO SHOT TTT++ CoTTA ActMAD TTT AdaContrast
Builds on: Tent SHOT AdaContrast MEMO CoTTA
Closest papers by meaning
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 2024 · similarity 0.89)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.88)
- Test-Time Adaptation with Binary Feedback (ICML 2025 · similarity 0.88)
- Multi-Label Test-Time Adaptation with Bound Entropy Minimization (ICLR 2025 · similarity 0.87)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · 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.