Home About Downloads Scheme
 
TEDXPeople TEDXPeople TEDXPeople
A Dataset of YouTube TED videos A Dataset of YouTube TED videos A Dataset of YouTube TED videos
 
 
 
 
 
 
Figure 1: TEDXPeople samples, i.e pictures of people performing on a stage during TED/TEDx talks.
 
 
About dataset
 
 
TEDXPeople dataset contains 48,188 videos of TED and TED-X talks. We have extracted and processed from 8 to 16 frames from each video, in which a person is visible in either full-body or close-up upper-body view. This dataset is very diverse in terms of demographics andclothing styles, though the resolution is limited and there are certain biases associated with peculiar lighting and camera viewpoints specific to TED(-X) talks. We open-source a list of links to YouTube videos and specific frames that we used. For each frame we also provide an SMPL-X fit of a speaker.
 
 
 
 
 
 
Figure 2: Full-body samples from a random TED video with corresponding SMPL-X fits.
 
 
Downloads
 
 
Due to legal reasons we can not publish raw images from the dataset. However, we provide list of youtube videos we used along with frame indexes used for training and testing. We also publish parameters for SMPLx body model fitted using smplify-x mathod.
 
You may download a .zip file with the data splits and SMPLx parameters using the following links:
 

Data splits

SMPLx Parameters
 
 
The data splits .zip file contains following data:
  • TEDXPeople.json -- full list of video links and frame indexes splitted into train and test parts
  • TEDXPeople_upperbody.json -- subset containing only upperbody frames
  • TEDXPeople_512.json -- subset containing only frames where human body resolution is close to or exceeds 512p
 
.json files scheme
 
 
 
 
 
 
To get the data you need to download youtube videos with the identifiers stored in .json files and extract corresponding frames from them.