Night Photography Rendering Challenge 2024 Dataset

The procedure of capturing and processing images taken by cameras involves employing onboard processing to transform raw sensor images into the final, polished photographs, subsequently encoding them in a standard color space like sRGB.

However, capturing images at night presents distinct challenges not typically encountered in daylight shots. Unlike daytime images, where assuming a single global illumination is often adequate, night images frequently feature multiple illuminants, many of which are visible in the scene. This complexity makes it difficult to determine the optimal illumination correction for rendering night images. Additionally, commonly used tone curves and photo-finishing techniques for daytime images may not be suitable for night photography. Furthermore, widely employed image metrics like SSIM and LPIPS may not effectively evaluate night images. The absence of established best practices and limited research in the realm of night photography pose significant challenges.

The main objective of addressing this challenge is to stimulate research and advance the field of image processing specifically tailored for night photography.

The participants in this challenge will be granted access to raw-RGB images of night scenes, which have been captured using the Huawei Mate 40 Pro sensor type and are encoded in 16-bit PNG files. Accompanying these images will be additional meta-data, provided in JSON files.  As an added resource, the organizers have also provided code for a baseline algorithm, as well as a demonstration, on GitHub.

Dataset includes data postprocessed by baseline algorithm and resulting images processed by competitors` algorithms.

Mean opinion scores are obtained through visual comparison, carried out using Yandex Tasks (similar to Mechanical Turk). Yandex Tasks users ranked the solutions in a forced-choice manner. It is important to note that Yandex Tasks primarily relies on observers from Eastern Europe and and Central Asia to perform the image ranking, and as a result, there may be a cultural bias in terms of the preferred image aesthetics by the observers. Yandex Tasks users were not aware of the identity of the participants. An example of this evaluation process can be found at a specified link.