Dai Yuekun ; Li Chongyi ; Zhou Shangchen ; Feng Ruicheng ; Loy Chen Change

Abstract

Artificial lights commonly leave strong lens flare artifacts on imagescaptured at night. Nighttime flare not only affects the visual quality but alsodegrades the performance of vision algorithms. Existing flare removal methodsmainly focus on removing daytime flares and fail in nighttime. Nighttime flareremoval is challenging because of the unique luminance and spectrum ofartificial lights and the diverse patterns and image degradation of the flarescaptured at night. The scarcity of nighttime flare removal datasets limits theresearch on this crucial task. In this paper, we introduce, Flare7K, the firstnighttime flare removal dataset, which is generated based on the observationand statistics of real-world nighttime lens flares. It offers 5,000 scatteringand 2,000 reflective flare images, consisting of 25 types of scattering flaresand 10 types of reflective flares. The 7,000 flare patterns can be randomlyadded to flare-free images, forming the flare-corrupted and flare-free imagepairs. With the paired data, we can train deep models to restoreflare-corrupted images taken in the real world effectively. Apart from abundantflare patterns, we also provide rich annotations, including the labeling oflight source, glare with shimmer, reflective flare, and streak, which arecommonly absent from existing datasets. Hence, our dataset can facilitate newwork in nighttime flare removal and more fine-grained analysis of flarepatterns. Extensive experiments show that our dataset adds diversity toexisting flare datasets and pushes the frontier of nighttime flare removal.