NTIRE 2026 第三届任意图像恢复模型 (RAIM) 挑战：AI 闪回人像挑战（第3轨道）
计算机视觉与模式识别
2026-04-14 v1
摘要
本文提供了关于 NTIRE 2026 第三届任意图像恢复模型 (RAIM) 挑战的全面概述，具体聚焦于第3轨道：AI 闪回人像。尽管深度学习在图像恢复方面取得了显著进展，但现有模型在现实世界低光人像场景中仍面临重大挑战。具体而言，它们难以在降噪、细节保留以及准确照明和色彩再现之间取得最佳平衡。为弥合这一差距，本次挑战旨�为现实世界低光人像恢复建立新基准。我们采用融合客观量化指标与严格主观评估协议的混合评估体系，对所提出的算法进行全面评估。本次比赛提供了包含800组真实拍摄低光人像数据的数据集。每组数据包含一张1K分辨率的低光输入图像、一张1K分辨率的 ground truth (GT) 以及一张1K分辨率的人脸掩码。本次挑战在学术界和工业界广受关注，吸引了超过100支参赛团队，获得了3000多次有效提交。本报告详细介绍了挑战的动机、数据集构建过程、评估指标以及比赛的各个阶段。该轨道发布的数据集和基线代码已在相同的 GitHub 仓库中公开，官方挑战网页 hosted 在 CodaBench 上。
引用
@article{arxiv.2604.11230,
title = {NTIRE 2026 The 3rd Restore Any Image Model (RAIM) Challenge: AI Flash Portrait (Track 3)},
author = {Ya-nan Guan and Shaonan Zhang and Hang Guo and Yawen Wang and Xinying Fan and Tianqu Zhuang and Jie Liang and Hui Zeng and Guanyi Qin and Lishen Qu and Tao Dai and Shu-Tao Xia and Lei Zhang and Radu Timofte and Bin Chen and Yuanbo Zhou and Hongwei Wang and Qinquan Gao and Tong Tong and Yanxin Qian and Lizhao You and Jingru Cong and Lei Xiong and Shuyuan Zhu and Zhi-Qiang Zhong and Kan Lv and Yang Yang and Kailing Tang and Minjian Zhang and Zhipei Lei and Zhe Xu and Liwen Zhang and Dingyong Gou and Yanlin Wu and Cong Li and Xiaohui Cui and Jiajia Liu and Guoyi Xu and Yaoxin Jiang and Yaokun Shi and Jiachen Tu and Liqing Wang and Shihang Li and Bo Zhang and Biao Wang and Haiming Xu and Xiang Long and Xurui Liao and Yanqiao Zhai and Haozhe Li and Shijun Shi and Jiangning Zhang and Yong Liu and Kai Hu and Jing Xu and Xianfang Zeng and Yuyang Liu and Minchen Wei},
journal= {arXiv preprint arXiv:2604.11230},
year = {2026}
}
备注
Accepted to CVPR 2026 Workshop. Includes supplementary material as ancillary file