You Described, We Archived:
A Rich Audio Description Dataset
Charity Pitcher-Cooper, Manali Seth, Benjamin Kao
James M. Coughlan, Ilmi Yoon
Happy Birthday to YouDescribe!
Launched in 2013, Rebuilt in 2017, Turning 10 in 2023
2
What is YouDescribe?
3
What YouDescribe isn’t
4
YouDescribed, We Archived (YuWA)
What’s a data paper and why write one?
5
What data do we collect?
(and don’t collect)
6
The WHO of the YuWA data set
Our viewers are diverse
YouDescribe is first and foremost an accessibility tool for the BVI community. (but lots of other groups use it too)
Videos can be viewed from anywhere in the world with internet access and a computer
We also have an iOS app for apple phones and tablets.
Our describers are diverse
Over 3,000 volunteer describers
Over 7,000+ audio descriptions
Approximately 1,000+ ADs added in 2022, and on average 90+ videos described each month
More volunteers providing descriptions each year
7
YouDescribe.org has 12,000+ average annual visitors
They are global- But mostly the US.
Number of audio descriptions added yearly
YEAR- # of descriptions
2017 - 287
2018 - 579
2019 - 718
2020 - 1104
2021 - 1324
8
The Magic of Premium Data Sets
One video- so many descriptions
Over 80 descriptions for a single 21 second video
One describer club, 5 years of videos
HS age students, with identical training
9
Is the data any good?
Oxford University Group QuerYD
Dataset trainer for retrieval and event localization in videos
Compared to other training data sets content descriptions made with YouDescribe data:
Our YuWA Dataset
Our data set is better
We have more metadata
We have twice the audio (both the video dialog and the audio descriptions)
We have extended descriptions
Better metadata means more research applications
10
You Described, We Archived Data Repository
URL : https://github.com/youdescribe-sfsu/You-Described-We-Archived
Steps to download the data
You Described, We Archived ©️ 2022 by Smith-Kettlewell Eye Research Institute, San Francisco State University is licensed under CC BY-NC-ND 4.0.
11
Possible AI/ML applications: AI-Enhanced Video Accessibility for YouDescribe
Ilmi Yoon - Professor, Dept. Computer Science, San Francisco State University
12
Data Statistics I
3100+ describers and
7,000+ audio descriptions (82,000+ audio clips ).
About 22% of the audio descriptions are rated out of which
59% are excellent,
25% are very good,
8% are good,
5% are fair, and
only 3% are rated as poor.
13
Video Categories
14
Data Statistics II – a typical YouDescribe video
Total Videos with High Quality AD: ~6300
Average duration of videos: 1,870,000 sec / 6300 => ~ 300 seconds ~5 minutes
Total audio descriptions: ~7000
Average duration of audio descriptions: ~7000000 sec / 7000 ~ => 100 seconds (or 1 minute 40 sec)
Total audio clips: ~82000
Average number of audio clips: 82000 / 7000 => ~ 12
Average duration of audio clips: ~ 8.5 seconds
Max duration of audio clip: 405.36 seconds ~ 6.75 minutes
Conclusion: 12 audio clips per description over 5 min videos, so about one audio clip per 25 seconds
15
Data Statistics III
16
Data Statistics IV
17
Multiple ADs per video
(total duration and counts of audio clips)
18
Multiple ADs per video
(Insertion points and description style decision)
19
Thank you for your attention!
A special thank you to YouDescribe inventor and mentor Dr Joshua Miele. There is no YouDescribe without you.
This work was supported by grants from:
20