Abstract
Artificial lights commonly leave strong lens flare artifacts on the images when one captures images at night. Nighttime flare not only affects the visual quality but also degrades the performance of vision algorithms. Different from sunshine flare, the nighttime flare has its unique luminance and spectrum of artificial lights and diverse patterns. Models trained on existing sunshine flare removal datasets, however, cannot cope with nighttime flare. Existing flare removal methods mainly focus on the removal of daytime flares while they fail in removing nighttime flares. Nighttime flare removal is challenging because of the unique luminance and spectrum of artificial lights and the diverse patterns and image degradation of the flares captured at night. The scarcity of the nighttime flare removal dataset limits the research on this paramount task. In this paper, we introduce, Flare7K, the first nighttime flare removal dataset, which is generated based on the observation and statistic of real-world nighttime lens flares. It offers 5,000 scattering flare images and 2,000 reflective flare images, consisting of 25 types of scattering flares and 10 types of reflective flares. The 7,000 flare patterns can be randomly added to the flare-free images, forming the flare-corrupted and flare-free image pairs. With the paired data, deep models can effectively restore the flare-corrupted images taken in real world. Apart from sufficient flare patterns, we also provide rich annotations, including the light source, glare with shimmer, reflective flare, and streak, which are frequently absent from existing datasets. Thus, our dataset can facilitate new work in nighttime flare removal and more. Extensive experiments demonstrate that our dataset can complement the diversity of existing flare datasets and push the frontier of nighttime flare removal.
Video
Flare Removal Result on Real Data

MouseOver: Nighttime deflared images

MouseOut: Nighttime flare-corrupted images

We referred to the project page of Nerfies, AvatarCLIP and Text2Human when creating this project page.

This dataset is licensed under CC BY-NC-SA 4.0.