Research map / Detection and segmentation / Open-set and few-shot detection
Learning to Control Camera Exposure via Reinforcement Learning
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 DRL-AE, a SAC-trained MLP agent that jointly controls camera exposure time and gain from a vectorized intensity history, trained in a light-controlled darkroom with curriculum and spatial domain randomization.
Problem
Camera auto-exposure control in dynamic lighting needs rapid convergence, real-time processing on lightweight hardware, and good image quality; existing one-by-one or black-box joint methods converge slowly or flicker.
Approach
Formulates joint exposure time and gain control as a continuous-relative-action RL problem with a lightweight vectorized intensity-history state, a flicker- and image-attribute-aware reward, a static-to-dynamic lighting curriculum, and spatial domain randomization, enabling zero-shot real-world transfer.
Key results
- Converges within about 5 frames in the darkroom and 3-5 frames in real-world tests, versus 30 frames indoors and 10 outdoors for built-in AE.
- Detects 38% more SIFT features on average than built-in AE across 5355 image pairs from a driving vehicle.
- Inference takes about 1 ms on a Ryzen 5950x CPU, versus 108.7 ms for Shin et al. at 1600x1200.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRL-AE instead of the built-in AE controller | Camera exposure control · Real-world indoor and outdoor | frames to converge: 30 frames for indoors and 10 frames for outdoors; ours converges within 3-5 frames | Helped |
| Use DRL-AE instead of Shin et al. [21] | Camera exposure control · Exposure Control Dataset | frames to converge: three frames; about 30 frames | Helped |
| Use DRL-AE instead of built-in AE | SIFT feature extraction · Real-time driving, 5355 image pairs | average SIFT features: 1,306 vs 946; 38% more | Helped |
| Use DRL-AE instead of built-in AE | SIFT feature extraction · Real-time driving, 5355 image pairs | median SIFT features: 1,157 vs 711; 62% | Helped |
| Use DRL-AE instead of built-in AE | YOLO-v5 object detection · Real-time driving | number of detected objects: 5% more | Helped |
| Use DRL-AE instead of Shin et al. [21] | Camera exposure control · Ryzen 5950x CPU, 1600 x 1200 | processing time (ms): 108.7 -> 1.0 | Helped |
| Replace SAC with DDPG (ablation) | DRL-AE · Light-controlled darkroom | reward per frame: 1.61 -> 1.11 | Hurt |
| Replace SAC with TD3 (ablation) | DRL-AE · Light-controlled darkroom | reward per frame: 1.61 -> 1.03 | Hurt |
| Use vectorized intensity state instead of CNN state (ablation) | DRL-AE · Light-controlled darkroom | reward per frame: 0.85 -> 1.61 | Helped |
| Use relative actions instead of absolute actions (ablation) | DRL-AE · Light-controlled darkroom | reward per frame: 0.65 -> 1.61 | Helped |
| Add the flicker reward to the noise-only reward (ablation) | DRL-AE · Light-controlled darkroom | frames to converge: 9 -> 5 | Helped |
| Add the noise reward to the flicker-only reward (ablation) | DRL-AE · Light-controlled darkroom | frames to converge: - -> 5 | Helped |
Surprising findings
- CNN feature-based state performed poorly (reward 0.85, never converged) compared with the simple vectorized intensity state.
Limitations
- Darkroom training environment has limited object context and no camera or object motion, so motion blur is not modeled.
- Aperture is not controlled; only exposure time and gain.
- Reward is a primitive brightness/flicker/noise formulation without modern image quality metrics.
Experimental setup
Datasets: Light-controlled darkroom (self-collected) Exposure control dataset (Shin et al.)
Benchmarks: Light-controlled darkroom test scenarios Exposure control dataset (Shin et al.) Real-world indoor/outdoor driving sequences
Metrics: reward per frame frames to converge processing time (ms) SIFT feature count detected object count
Baselines: Camera built-in auto-exposure Shin et al. Nelder-Mead exposure control Tomasi et al. exposure network DDPG TD3 CNN state design
Builds on: Soft Actor-Critic TD3 DDPG PPO
Closest papers by meaning
- Goal Conditioned Reinforcement Learning for Photo Finishing Tuning (NeurIPS 2024 · similarity 0.86)
- RL-SeqISP: Reinforcement Learning-Based Sequential Optimization for Image Signal Processing (AAAI 2024 · similarity 0.83)
- A Simple Framework for Generalization in Visual RL under Dynamic Scene Perturbations (NeurIPS 2024 · similarity 0.82)
- AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers (CVPR 2025 · similarity 0.82)
- Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning (CVPR 2025 · similarity 0.82)
- VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.82)
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.