LoVR
收藏资源简介:
LoVR是一个专为长视频文本检索设计的基准数据集,由华东师范大学、北京大学等研究机构创建。数据集包含467个完整长视频和超过40,804个细粒度剪辑,每个视频和剪辑都配有高质量的文本描述。LoVR旨在支持全视频和细粒度剪辑级别的检索,并克服了现有基准数据集在视频时长、字幕质量和标注粒度方面的限制。数据集的创建过程包括视频选择、剪辑分割、字幕生成和全视频字幕构建等步骤。LoVR的应用领域为视频理解和检索,旨在解决现有方法在处理长视频时的局限性。
LoVR is a benchmark dataset specifically designed for long-video text retrieval, developed by research institutions including East China Normal University, Peking University, and other relevant organizations. The dataset comprises 467 full-length long videos and over 40,804 fine-grained clips, with each video and clip accompanied by high-quality textual descriptions. LoVR is intended to support retrieval tasks at both the full-video and fine-grained clip levels, and addresses the limitations of existing benchmark datasets regarding video duration, caption quality, and annotation granularity. The dataset creation workflow includes steps such as video selection, clip segmentation, caption generation, and full-video caption construction. The application domains of LoVR cover video understanding and retrieval, and it aims to mitigate the limitations of existing methods in long-video processing.

- 1LoVR: A Benchmark for Long Video Retrieval in Multimodal Contexts华东师范大学、北京大学、北京航空航天大学、北京理工大学、华中科技大学 · 2025年



