MomentSeeker
收藏资源简介:
我们提出了MomentSeeker,一个全面的基准测试,用于评估检索模型在处理一般长视频时刻检索(LVMR)任务中的性能。MomentSeeker具有三个关键优势。首先,它包含了平均超过500秒的长视频,使其成为首个专门用于长视频时刻检索的基准测试。其次,它涵盖了广泛的任务类别(包括时刻搜索、字幕对齐、图像条件时刻搜索和视频条件时刻搜索)和多样化的应用场景(如体育、电影、卡通和自我),使其成为评估检索模型一般LVMR性能的全面工具。此外,评估任务经过人工精心策划,确保评估的可靠性。我们进一步在合成数据上微调了一个基于MLLM的LVMR检索器,该检索器在我们的基准测试中表现出强大的性能。检查点将很快发布。
We propose MomentSeeker, a comprehensive benchmark for evaluating the performance of retrieval models in handling the Long Video Moment Retrieval (LVMR) task. MomentSeeker boasts three key advantages. Firstly, it includes long videos averaging over 500 seconds, making it the first benchmark specifically designed for long video moment retrieval. Secondly, it covers a wide range of task categories (including moment search, subtitle alignment, image-based moment search, and video-based moment search) and diverse application scenarios (such as sports, movies, cartoons, and self), positioning it as a comprehensive tool for assessing the general LVMR performance of retrieval models. Furthermore, the evaluation tasks have been meticulously planned by humans to ensure the reliability of the assessment. We have also fine-tuned a MLMM-based LVMR retriever on synthetic data, which demonstrated strong performance in our benchmark. Checkpoints will be released soon.
MomentSeeker数据集概述
基本信息
- 数据集名称: MomentSeeker
- 论文标题: MomentSeeker: A Comprehensive Benchmark and A Strong Baseline For Moment Retrieval Within Long Videos
- 论文链接: https://arxiv.org/abs/2502.12558
- 数据集链接: https://huggingface.co/datasets/avery00/MomentSeeker
- 许可证: CC-BY-NC-SA-4.0
数据集特点
- 视频长度: 平均超过500秒的长视频
- 任务类别:
- Moment Search
- Caption Alignment
- Image-conditioned Moment Search
- Video-conditioned Moment Search
- 应用场景: 体育、电影、卡通、ego等多样化场景
- 标注方式: 人工标注确保评估可靠性
评估与基准
- 评估指标: Recall@1, MAP@5
- 评估方法: 提供JSON文件包含每个问题的候选视频,可进行排序和指标计算
- 基准模型: V-Embedder (InternVideo2-Chat, 8B参数) 表现最佳
使用限制
- 用途限制: 仅限研究使用,禁止商业用途
- 版权声明: 不拥有原始视频文件的版权,已对视频进行降分辨率、剪辑等处理
- 移除请求: 如原视频作者要求移除,将替换为稀疏采样的视频帧或元信息
维护计划
- 长期保留: 标注文件将永久保留
- 视频替代方案: 如视频被移除,将替换为稀疏采样的视频帧或元信息
引用信息
bibtex @misc{yuan2025momentseekercomprehensivebenchmarkstrong, title={MomentSeeker: A Comprehensive Benchmark and A Strong Baseline For Moment Retrieval Within Long Videos}, author={Huaying Yuan and Jian Ni and Yueze Wang and Junjie Zhou and Zhengyang Liang and Zheng Liu and Zhao Cao and Zhicheng Dou and Ji-Rong Wen}, year={2025}, eprint={2502.12558}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2502.12558}, }




