多源基准与眼自适应像素MeanFlow去除眼镜反光
原标题：Seeing Through the Glare: A Multi-Source Benchmark and Ocular-Adaptive Pixel MeanFlow for Eyeglass Reflection Removal
AI 摘要
该论文提出 OcuBench 多源基准（含 10,280 合成对、732 真实输入伪对、458 真实测试图）与 OcuFlow 框架，用于眼镜反光去除。OcuFlow 基于几何自适应 token 支持与几何引导的 pMF，实现单次前向生成，在 RTX 4090 上每图 56 ms，并在配对基准、真实场景无参考评估、用户偏好及下游虹膜定位上均优于现有方法。
正文节选
19365b \setthemecolor1a2a6c \titlelogocontent \settitlepreskip-10ex \settitlelogogap3em \showlogointitleOcuFlow Seeing Through the Glare: A Multi-Source Benchmark and Ocular-Adaptive Pixel MeanFlow for Eyeglass Reflection Removal 1 Introduction Reflections from windows or monitors on eyeglasses can obscure the eyes while leaving most of a portrait intact. Although spatially confined, this interference can noticeably change facial appearance in photographs and video communication. Its effects ran