24,330 paired images featuring underexposed and overexposed photographs alongside expert-retouched ground truth versions. The dataset includes images rendered at five different exposure levels to simulate common photographic errors and facilitate multi-scale image restoration.
Use Cases
- Train image-to-image translation models using the underexposed/overexposed inputs and ground_truth pairs
- Develop exposure correction algorithms that specifically target detail recovery in saturated overexposed pixels
- Benchmark the performance of multi-scale neural networks on high-resolution photo enhancement tasks
Strengths
- 24,330 image pairs derived from the MIT-Adobe FiveK dataset
- Includes five discrete exposure levels ranging from -1.5 to +1.5 EV
- Features expert-retouched ground truth images for high-quality color and contrast targets
- Supports multi-scale training architectures to address both local and global exposure artifacts