TED-talks

为了创建 TED-talks 数据集，使用“TEDtalks”查询下载了 3,035 个 YouTube 视频。从这些初始候选者中，选择了人物上半部分可见至少 64 帧且人物边界框高度至少为 384 像素的视频。静态视频被手动过滤掉，其中一个人正在做其他事情而不是展示的视频。

To construct the TED-talks dataset, 3,035 YouTube videos were downloaded using the search query "TEDtalks". From these initial candidate videos, selection was conducted based on two criteria: the upper body of the speaker is visible for a minimum of 64 frames, and the height of the speaker's bounding box is at least 384 pixels. Static videos and clips where the speaker was engaged in other activities instead of delivering a talk were manually filtered out.

搜集汇总

背景与挑战

背景概述

TED-talks数据集是一个用于视频重构的公开数据集，包含3,035个从YouTube下载的TED演讲视频，经过筛选以确保人物上半部分可见且边界框高度足够。该数据集由特伦托大学于2021年发布，采用CC BY-NC-ND 4.0许可，相关论文可在arXiv上获取。