A Real Low-Light Image Super-Resolution Dataset

Andreas Aakerberg, Kamal Nasrollahi, Thomas B. Moeslund

NeurIPS 2021

Paper | Dataset|


Digital images can suffer from several different degradations that reduce the visibility and level of details in the images. These degradations can occur both due to environmental factors in the scene, and limitations of the hardware. Two common degradation types are under-exposure, due to poor illumination of the scene, and low resolution, due to the limited spatial resolution of the image sensor. However, the reconstruction of images with these degradations is currently only being addressed separately in the literature. The RELLISUR dataset aims to fill the gap between low-light image enhancement and low-resolution image enhancement (Super-Resolution (SR)) by providing a large set of real low-light low-resolution images paired with normal-light high-resolution reference image counterparts. In total the dataset contains 12750 paired images of different resolutions and degrees of low-light illumination, to facilitate learning of deep-learning-based models that can perform a direct mapping from degraded images with low visibility to high-quality detail-rich images of high resolution.


The RELLISUR dataset is the first dataset that contains real low-light low-resolution and normal-light high-resolution image pairs. The dataset is made publicly available and contains a large number of in- and outdoor scenes captured by a DSLR camera. The dataset is a collection of image sequences containing real ×1, ×2, and ×4 normal-light images, together with five real low-light images. The ×1 and ×2 scale levels represent the low-resolution images while the ×4 scale level represents the high-resolution Ground-Truth reference images. The low-light images are acquired at scale ×1 and are also considered low-resolution. In total, the RELLISUR dataset consists of 850 distinct sequences. An example of a sequence can be seen in the figure above. With three different scale levels, the total number of normal light low-resolution and high-resolution pairs is 2550. As the five under-exposed images in a sequence corresponds to the same normal-light reference image, the resulting number of low-light / normal-light image pairs is 4250 for each of the three scale levels. Hence, the total number of low-light / normal-light images pairs in the RELLISUR is 12750. Please refer to our paper for more details.


Creative Commons Attribution 4.0 International


@inproceedings{ aakerberg2021rellisur, 
title={{RELLISUR}: A Real Low-Light Image Super-Resolution Dataset}, 
author={Andreas Aakerberg and Kamal Nasrollahi and Thomas B. Moeslund}, 
booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)}, 
url={} }