Research map / Label-efficient and robust learning / Test-time adaptation
Adaptive Camera Sensor for Vision Models
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 Lens, a training-free test-time camera sensor control method that selects sensor parameters by model confidence (VisiT), and introduces the ImageNet-ES Diverse benchmark of physically captured images under varied sensor and lighting conditions.
Problem
Domain shift from low-quality, covariate-shifted images degrades vision model accuracy; existing fixes modify the model, while human-centric auto-exposure does not optimize for model perception.
Approach
Select camera sensor parameters per model and scene at test time by scoring captured images with the target model's own softmax confidence (VisiT), with candidate selection algorithms limiting the number of captures to keep latency low, leaving the model unchanged.
Key results
- Lens with VisiT improves accuracy over auto-exposure and random sensor control across all models, with gains of 8.71% to 47.58%.
- Lens on EfficientNet-B3 (12M params) outperforms DG-enhanced OpenCLIP-h (632M params) by 7.34% accuracy.
- CSA3 cuts image capture time by 93.4% (to 0.16 s) while keeping accuracy close to full 27-option search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply model- and scene-specific sensor control with VisiT instead of auto-exposure | Vision models across the evaluated benchmarks · ImageNet-ES | accuracy: 59.8 -> 85.6 | Helped |
| Apply model- and scene-specific sensor control with VisiT instead of auto-exposure | Vision models across the evaluated benchmarks · ImageNet-ES Diverse | accuracy: 34.2 -> 59.8 | Helped |
| Use model- and scene-specific sensor control rather than fixed sensor parameters | Vision models across the evaluated benchmarks · ImageNet-ES | accuracy: 73.4 -> 95.4 | Helped |
| Replace auto-exposure data with Lens-acquired data | ResNet-18 · ImageNet-ES | accuracy: 39.2 -> 73.8 | Helped |
| Replace auto-exposure data with Lens-acquired data | ResNet-18 · ImageNet-ES Diverse | accuracy: 13.1 -> 34.6 | Helped |
| Replace VisiT confidence with an OOD score for sensor candidate selection (ablation) | Lens with ResNet-18 · ImageNet-ES | accuracy: 53.4 -> 73.8 | Hurt |
| Replace VisiT confidence with entropy of logits (ablation) | Lens with ResNet-50 · ImageNet-ES Luminous | accuracy: 78.6 -> 78.8 | No clear effect |
| Replace VisiT confidence with entropy of logits (ablation) | Lens with ResNet-50 · ImageNet-ES Diverse | accuracy: 43.3 -> 42.9 | Hurt |
| Use cost-based candidate selection with K=18 instead of all 27 sensor options (ablation) | Lens with EfficientNet-B0 · ImageNet-ES | accuracy: 77.8 -> 78.6 | Helped |
| Use cost-based candidate selection with K=18 instead of all 27 sensor options (ablation) | Lens with EfficientNet-B0 · ImageNet-ES Diverse | accuracy: 39.6 -> 31.4 | Hurt |
Surprising findings
- Images that look worse to humans (dark or bright) selected by Lens yield better model accuracy than human-centered auto-exposure settings.
- OOD scores (ViM, ReAct, ASH, KNN) are less reliable than confidence as quality proxies under covariate shift.
Limitations
- Model confidence may be overconfident, especially for poorly calibrated models.
- Only classification evaluated in main experiments; other tasks only qualitatively analyzed in appendix.
- Sensor control limited to discrete manual parameter options (27 settings) and does not address radial distortion or framing.
Experimental setup
Datasets: ImageNet-ES Tiny ImageNet ImageNet-ES Diverse
Benchmarks: ImageNet-ES ImageNet-ES Diverse
Metrics: top-1 accuracy capture latency (sec)
Baselines: Auto-Exposure Random sensor selection BN1 BN2 TENT DeepAugment+AugMix ViM ReAct ASH KNN
Builds on: ImageNet-ES Tiny ImageNet test-time adaptation (TENT) DeepAugment AugMix
Closest papers by meaning
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.83)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.83)
- Adaptive Parameter Selection for Tuning Vision-Language Models (CVPR 2025 · similarity 0.83)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.83)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.83)
- Seeing More with Less: Human-like Representations in Vision Models (CVPR 2025 · similarity 0.83)
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.