NTIRE 2026 第三届任意图像恢复模型挑战赛：专业图像质量评估（赛道一）
计算机视觉与模式识别
2026-04-15 v1 人工智能
摘要
本文概述了NTIRE 2026第三届野外任意图像恢复模型挑战赛，特别聚焦于赛道一：专业图像质量评估。传统的图像质量评估通常依赖标量分数。通过将复杂的视觉特征压缩为单一数值，这些方法从根本上难以区分均匀高质量图像间的细微差异。此外，它们无法阐明为何一幅图像更优，缺乏为视觉任务提供指导所需的推理能力。为弥补这一差距，多模态大语言模型的最新进展提供了一种有前景的范式。受此潜力启发，本挑战赛建立了一个新颖的基准，探索MLLM在评估高质量图像对时模仿人类专家认知的能力。参赛者需克服专业场景中的关键瓶颈，聚焦于两个主要目标：(1) 比较质量选择：可靠地识别高质量图像对中视觉上更优的图像；(2) 解释性推理：生成有依据的、专家级别的解释，详细说明选择背后的理由。该挑战赛共吸引了近200个注册和超过2500次提交。表现最佳的方法显著推进了专业IQA领域的最新技术水平。挑战数据集可在https://github.com/narthchin/RAIM-PIQA获取，官方主页为https://www.codabench.org/competitions/12789/。
引用
@article{arxiv.2604.12512,
title = {NTIRE 2026 The 3rd Restore Any Image Model (RAIM) Challenge: Professional Image Quality Assessment (Track 1)},
author = {Guanyi Qin and Jie Liang and Bingbing Zhang and Lishen Qu and Ya-nan Guan and Hui Zeng and Lei Zhang and Radu Timofte and Jianhui Sun and Xinli Yue and Tao Shao and Huan Hou and Wenjie Liao and Shuhao Han and Jieyu Yuan and Chunle Guo and Chongyi Li and Zewen Chen and Yunze Liu and Jian Guo and Juan Wang and Yun Zeng and Bing Li and Weiming Hu and Hesong Li and Dehua Liu and Xinjie Zhang and Qiang Li and Li Yan and Wei Dong and Qingsen Yan and Xingcan Li and Shenglong Zhou and Manjiang Yin and Yinxiang Zhang and Hongbo Wang and Jikai Xu and Zhaohui Fan and Dandan Zhu and Wei Sun and Weixia Zhang and Kun Zhu and Nana Zhang and Kaiwei Zhang and Qianqian Zhang and Zhihan Zhang and William Gordon and Linwei Wu and Jiachen Tu and Guoyi Xu and Yaoxin Jiang and Cici Liu and Yaokun Shi},
journal= {arXiv preprint arXiv:2604.12512},
year = {2026}
}
备注
NTIRE Challenge Report. Accepted by CVPRW 2026