The challenge on
HDR Reconstruction from a Single Raw Image

Data Collection

We capture a real paired Raw-to-HDR dataset for HDR reconstruction from a single Raw image. The captured dataset covers a large range of HDR scenarios, including modern/ancient buildings, art districts, tourist attractions, street shops and restaurants, abandoned factories, city views and so on. Those images are captured at different times of the day, including daytime and nighttime, which further guarantees the diversity of the paired Raw-to-HDR dataset. The data capture process involves several steps. Initially, we carefully select scenes with high dynamic range potential. Then, using a Canon 5D Mark IV camera mounted on a tripod, we employ bracket exposure mode to capture different exposures of the same scene. The raw images taken in challenging lighting conditions, specifically from -3EV to +3EV, are used as input images. The corresponding ground truth images are created using an HDR merging method, as described by Debevec et al. (2008).

Dataset characteristics

High resolution
Our dataset stands out with its high resolution of 6720x4480, surpassing the common resolutions (below 1920x1080) found in other HDR datasets. This higher resolution captures finer details, offering a more comprehensive analysis for HDR reconstruction.

High bit-depth ground truth
The SRHDR dataset features ground truth HDR images with a bit-depth of over 20 bits, utilizing a linear HDR format. This high bit-depth ensures a richer and more precise representation of color and light intensities.

Real paired samples
Each image pair in the dataset is meticulously captured through multi-exposure fusion. The input comprises actual images shot with a DSLR under challenging lighting conditions. The corresponding ground truth HDR images are generated using a widely accepted HDR merging algorithm, ensuring authenticity and relevance.

Raw images as input
The use of unprocessed raw sensor data as the input format leverages the higher bit-depth and superior intensity tolerance of raw data, effectively addressing the common issue of insufficient scene information in HDR image processing.