LSVOS 2025 挑战报告：复杂视频对象分割的最新进展
计算机视觉与模式识别
2025-10-14 v1
摘要
本报告概述了于 ICCV 2025 上举办的第七届大规模视频对象分割挑战赛（LSVOS）。除面向真实视频场景中鲁棒性的两大传统赛道（经典视频对象分割 VOS 与指代视频对象分割 RVOS）外，2025 年版度新引入了复杂视频对象分割（MOSEv2）赛道。基于先前的洞见，MOSEv2 大幅提升了难度，引入了更具挑战性但更贴近现实的场景，包括更密集的小目标、频繁的消失/复现事件、严重遮挡、恶劣天气与光照等情况，推动了在精心构建的基准测试之外的长期一致性与泛化能力。挑战赛保留了 VOS 与 RVOS 的标准 、 与 指标，而 MOSEv2 采用 作为主要排名指标，以更好地评估跨尺度对象及消失情况。我们总结了数据集与协议，概述了顶级方案，提炼出新兴趋势，如 LLM/MLLM 组件在增长作用以及记忆感知的传播，旨在为野外场景中具备鲁棒性与语言感知的视频分割指明未来方向。
引用
@article{arxiv.2510.11063,
title = {LSVOS 2025 Challenge Report: Recent Advances in Complex Video Object Segmentation},
author = {Chang Liu and Henghui Ding and Kaining Ying and Lingyi Hong and Ning Xu and Linjie Yang and Yuchen Fan and Mingqi Gao and Jingkun Chen and Yunqi Miao and Gengshen Wu and Zhijin Qin and Jungong Han and Zhixiong Zhang and Shuangrui Ding and Xiaoyi Dong and Yuhang Zang and Yuhang Cao and Jiaqi Wang and Chang Soo Lim and Joonyoung Moon and Donghyeon Cho and Tingmin Li and Yixuan Li and Yang Yang and An Yan and Leilei Cao and Feng Lu and Ran Hong and Youhai Jiang and Fengjie Zhu and Yujie Xie and Hongyang Zhang and Zhihui Liu and Shihai Ruan and Quanzhu Niu and Dengxian Gong and Shihao Chen and Tao Zhang and Yikang Zhou and Haobo Yuan and Lu Qi and Xiangtai Li and Shunping Ji and Ran Hong and Feng Lu and Leilei Cao and An Yan and Alexey Nekrasov and Ali Athar and Daan de Geus and Alexander Hermans and Bastian Leibe},
journal= {arXiv preprint arXiv:2510.11063},
year = {2025}
}
备注
16 pages, 9 figures