Seeing Through Flare: A Real-world Event-guided Benchmark for Flare Removal in the Dark

RE-Flare has 7,680 images, including 1,820 triplets of flare-corrupted, flare-free, and low-light images captured by conventional cameras. Specifically, to tackle the challenge of accurately estimating flare location and shape, we additionally offer 1,110 pairs of light masks and light-source edge masks based on the event stream, providing diverse choices for data-driven implementations.

RE-Flare数据集包含7680张图像，其中包括1820组由传统相机拍摄的眩光污染图像、无眩光图像和低光图像三元组。具体而言，为应对准确估计眩光位置与形状这一挑战，我们还基于事件流提供了1110对光掩模与光源边缘掩模，为数据驱动的实现方案提供了多样化选择。