KWAI-AD-AudVis

该数据集名为KWAI-AD-AudVis，包含了来自中国热门短视频应用快手上的85,432个广告视频，总时长约913小时。这些视频根据行业进行了标注，并分为19个主题类别。值得注意的是，该数据集中的视频在音频和视觉流中存在信息不一致的情况，它们之间的对应关系并不明确，这使得它特别适合用于多模态学习和广告推荐任务。该数据集的规模为85,432个视频，总时长约913小时，其研究任务主要是音频视觉对应学习。

The dataset, named KWAI-AD-AudVis, contains 85,432 advertising videos sourced from Kuaishou, a popular short-video application in China, with an overall duration of approximately 913 hours. These videos are annotated based on industry sectors and categorized into 19 thematic classes. Notably, the videos in this dataset suffer from information inconsistency between their audio and visual streams, with unclear corresponding relationships between the two modalities, which renders it particularly suitable for multimodal learning and advertising recommendation tasks. Boasting a scale of 85,432 videos and a total duration of roughly 913 hours, the primary research tasks centered on this dataset revolve around audio-visual correspondence learning.

搜集汇总

背景与挑战

背景概述

KWAI-AD-AudVis是一个包含85,432个快手广告视频的多模态数据集，由商业广告主制作，模拟个人用户视频风格，具有质量控制。数据集特点包括视觉与音频信息不一致、音视频对应关系模糊以及不同行业间风格差异，适用于多模态学习和广告推荐研究。