HowTo100M

该数据集名为HowTo100M，是一个大规模的教学视频集合，其中包含了人类在日常活动中执行各种任务的短片段。尽管该数据集被用于预训练视频和语言模型（VLMs），但它并不包含机器人操作数据。作为一个大规模的数据集，其任务主要是用于预训练视频与语言结合的模型。

The dataset named HowTo100M is a large-scale collection of instructional videos, containing short video clips of humans performing diverse tasks during daily activities. Although this dataset has been utilized for pre-training Video-Language Models (VLMs), it does not include robotic manipulation data. As a large-scale dataset, its core application is the pre-training of video-language fusion models.

搜集汇总

背景与挑战

背景概述

HowTo100M是一个大规模的教学视频数据集，包含1.36亿个带字幕的视频片段，源自120万个YouTube视频，涵盖23,000种不同领域的活动，如烹饪、手工艺、个人护理等。每个视频都配有从YouTube自动下载的字幕旁白。