The challenge on
Raw Image Based Over-Exposure Correction

Data Collection

The RPO dataset is the first specialized collection to systematically study the generality and practicality of over-exposure correction models. The dataset encompasses both indoor and outdoor scenes, captured in daylight or under direct illumination to avoid flickering. Each short-exposure (normal-exposure) image is paired with long-exposure (over-exposure) images with 4 ratios (x3, x5, x8, x10). The use of mirrorless cameras like the Canon EOS 5D Mark IV, equipped with a full-frame CMOS sensor, ensured high-resolution captures and minimized vibrations.

Dataset Composition

Short Exposure Images (Normal, GT)
Captured in each scene using a tripod-mounted camera. The camera was set to automatic mode to find optimal aperture and exposure time settings, then switched to manual mode to lock these settings. Images were taken using a remote mobile app to control the shutter, minimizing lens vibration.

Long Exposure Images (Over-exposure, OE)
Following the capture of short exposure GT images, only the "exposure time" setting was adjusted using the mobile app to simulate real over-exposure caused by incorrect settings. Four predetermined over-exposure ratios were used (×3, ×5, ×8, ×10). It was ensured that the camera was not touched during both long and short exposure captures to prevent any misalignment due to lens vibration.