A Deep-RL framework for fast, robust camera auto-exposure control.
Learning-to-Expose trains a Deep Reinforcement Learning (DRL) agent that adjusts exposure in real time, producing well-exposed frames while avoiding abrupt brightness changes.
Key features:
| Feature | Description |
|---|---|
| Real-world data | 14,430 images captured in 130 real-life scenes |
| Multi-component reward | Balances perceptual image quality (PSNR/SSIM) with temporal smoothness |
| Discrete-action DQN | Efficient training (≈3 steps to converge per episode) |
| 12× faster | Compared with camera-in-the-loop RL alternatives |
| Promising OOD generalization | Trained indoors, tested on unseen outdoor scenes |
ExpoSweep underpins the project with diverse, exposure-swept image stacks.
| Split | Scenes | Images | Notes |
|---|---|---|---|
| Train | 100 indoor | 11,100 | Offices, kitchens, labs, libraries |
| Test-In | 15 indoor | 1,665 | Held-out indoor scenes for ID evaluation |
| Test-OOD | 15 outdoor | 1,665 | Held out outdoor scenes (Gardens, parking lots, sports courts) for OOD evaluation |
Each scene folder contains PNGs whose filenames encode the exposure value (EV).
Dataset is provided in ExpoSweep.
# Clone
git clone https://github.com/xi1ngang/Learning-to-Expose.git
cd Learning-to-Expose
# Create env
conda create -n DRL_AEC python=3.11
conda activate DRL_AEC
pip install -r requirements.txt