Twilight Cowboy Dataset
收藏资源简介:
该数据集是NTIRE 2026低光增强挑战赛(Twilight Cowboy)的核心资源,由竞赛组织者精心构建,旨在推动低光计算摄影研究。数据集包含585个真实场景,每个场景由5张手持拍摄的RAW图像与1张高质量参考图像组成,覆盖室内外夜间环境。创建时先通过三脚架拍摄270帧RAW影像,经强度对齐与像素级鲁棒均值估计生成无噪声参考,再手持拍摄并每隔十帧选取五幅输入,模拟自然手抖。该数据集专为多帧RAW图像对齐与去噪任务设计,旨在解决日常移动摄影中噪声与运动模糊的联合挑战。
This dataset serves as the core resource for the NTIRE 2026 Low-Light Enhancement Challenge (Twilight Cowboy), meticulously developed by the competition organizers to promote research in low-light computational photography. It comprises 585 real-world scenes, each containing 5 handheld RAW images and one high-quality reference image, covering both indoor and outdoor nighttime environments. During dataset construction, 270 RAW frames were first captured with a tripod, and noise-free reference images were generated through intensity alignment and pixel-wise robust mean estimation. Subsequently, 5 input images were selected at every 10-frame interval from the handheld capture sessions to simulate natural camera shake. This dataset is specifically tailored for multi-frame RAW image alignment and denoising tasks, with the goal of addressing the combined challenge of noise and motion blur in everyday mobile photography.




