VCSL（Video Copy Segment Localization）

我们介绍了VCSL (视频复制段本地化)，这是一个新的综合段级带注释的视频复制数据集。与受视频级注释或小规模限制的现有复制检测数据集相比，VCSL不仅具有多两个数量级的段级标记数据，160k个逼真的视频复制对包含超过280k个本地化复制段对，而且还涵盖了各种视频类别和广泛的视频时长。手动提取每个收集的视频对中的所有复制段，并附有精确注释的开始和结束时间戳。

We introduce VCSL (Video Copy Segment Localization), a novel comprehensive segment-level annotated video copy detection dataset. Unlike existing copy detection datasets that are constrained by either video-level annotations or limited scales, VCSL not only features segment-level labeled data with two orders of magnitude greater volume: it encompasses 160k realistic video copy pairs that collectively contain over 280k localized copy segment pairs, but also covers a diverse range of video categories and a broad spectrum of video durations. All copied segments within each collected video pair were manually extracted, accompanied by precisely annotated start and end timestamps.

搜集汇总

背景与挑战

背景概述

VCSL是一个用于视频复制段本地化的综合数据集，包含超过160k个真实视频复制对和280k个带精确时间戳的本地化段，覆盖广泛视频类别和时长。该数据集由蚂蚁集团于2022年6月发布，旨在提供大规模段级注释资源以支持相关研究。