Research map / Detection and segmentation / CNN features
Self-Bootstrapping for Versatile 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 SPA, a versatile fully test-time adaptation method that uses weak-to-strong self-bootstrapping with Fourier-domain geometry-preserving augmentations (low-frequency amplitude masking and high-frequency noise injection) to work across classification, regression, segmentation and 3D detection.
Problem
Existing test-time adaptation objectives are tied to classification (entropy, prototypes) or require source statistics (alignment), and consistency augmentations like random cropping/resizing destroy object geometry, so TTA fails on object- and pixel-level tasks such as detection and segmentation.
Approach
Self-bootstrapping TTA where the original test image's predictions serve as a stop-gradient target for a weak view, with the weak view created by masking low-frequency Fourier amplitudes and injecting Gaussian noise so that geometry is preserved while learning signals are maximized; adds confidence-based active selection for classification.
Key results
- On ImageNet-C (severity 5, ViT-Base), SPA reaches 70.1% average accuracy vs 67.8% for EATA and 66.3% for ActMAD; combined with ActMAD and TENT it reaches 71.6%.
- On KITTI-C 3D monocular detection, SPA improves over MonoTTA by 1.4% average AP_3D|R40 for Cars and 2.4% for Pedestrians on the reported split.
- On continual Cityscapes-to-ACDC segmentation with Segformer-B5, SPA alone gets 59.2% mIoU average over three rounds, and combined with CoTTA it raises CoTTA's 58.5% to 60.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPA instead of EATA | ViT-Base test-time adaptation · ImageNet-C severity 5 | accuracy: 67.8 -> 70.1 | Helped |
| Add SPA to ActMAD and TENT | ViT-Base test-time adaptation · ImageNet-C severity 5 | accuracy: 71.6 | Helped |
| Use SPA instead of MonoTTA | MonoFlex test-time adaptation · KITTI-C, Pedestrians average | AP3D|R40: 9.5 -> 11.9 | Helped |
| Add SPA to CoTTA | Segformer-B5 test-time adaptation · ACDC continual TTA, rounds 1–3 | mIoU: 58.5 -> 60.7 | Helped |
| Remove weak-to-strong learning (ablation) | SPA · ImageNet-C severity 5 | accuracy: 70.1 -> 4.9 | Hurt |
| Add active selection (ablation) | SPA · ImageNet-C severity 5 | accuracy: 69.4 -> 70.1 | Helped |
| Add low-frequency amplitude mask (ablation) | SPA · ImageNet-C severity 5 | accuracy: 65.6 -> 70.1 | Helped |
| Add high-frequency noise injection (ablation) | SPA · ImageNet-C severity 5 | accuracy: 67.7 -> 70.1 | Helped |
| Use high-frequency mask alone instead of low-frequency mask plus noise injection (ablation) | SPA augmentation · ImageNet-C severity 5 | accuracy: 70.1 -> 16.8 | Hurt |
| Use low-frequency mask plus noise injection rather than either augmentation alone (ablation) | SPA augmentation · ImageNet-C severity 5 | accuracy: 67.8 -> 70.1; 65.6 -> 70.1 | Helped |
| Use SPA's Fourier augmentations instead of Gaussian blur (ablation) | Self-bootstrapping TTA · ImageNet-C severity 5 | accuracy: 25.4 -> 70.1 | Helped |
| Use separate low-frequency mask and noise views rather than one combined view (ablation) | SPA · ImageNet-C severity 5 | accuracy: 69.0 -> 70.1 | Helped |
Surprising findings
- Masking only high-frequency amplitudes causes model collapse (4.9% accuracy when weak-to-strong is removed and 16.8% average for high-frequency mask alone), contrary to intuition that high-frequency masking would be a natural augmentation.
- Bidirectional consistency (removing stop-gradient) causes drastic degradation, dropping ImageNet-C accuracy to 4.9%.
Limitations
- Confidence-aware selection is only defined for classification heads; regression and dense tasks use other filtering (e.g., thresholds inherited from MonoTTA/CoTTA).
- Noise injection factor must be tuned per task; optimal gamma range is narrower for dense 3D detection than for classification.
Experimental setup
Datasets: ImageNet-C ImageNet-R ImageNet-A ImageNet-Sketch KITTI-C ACDC Cityscapes
Benchmarks: ImageNet-C ImageNet-R ImageNet-A ImageNet-Sketch KITTI-C ACDC
Metrics: accuracy AP_3D|R40 mIoU
Baselines: BN Adapt TENT EATA SAR DeYO ActMAD CoTTA DePT MonoTTA VDP
Builds on: BYOL DINO Tent CoTTA MonoTTA
Closest papers by meaning
- BoostAdapter: Improving Vision-Language Test-Time Adaptation via Regional Bootstrapping (NeurIPS 2024 · similarity 0.87)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.87)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.86)
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.86)
- Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM (CVPR 2025 · similarity 0.86)
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.