2017-01-23T14:43:28 n9197

SAIVT-BNEWS

Viewed: 493

SAIVT-BNEWS

Overview

The SAIVT-BNews database consists of multi-modal annotation for a corpus of 55 Australian broadcast news videos. For each video, medadata, speech and speaker ground truth, face timing and identity ground truth, face locations, and an on screen text transcription are provided. The videos are not included within the archive, however a script to automatically download them is provided. Contact Dr David Dean or Dr Simon Denman for further information.


This distribution contains the SAIVT-BNEWS database, consisting of ground truth information and metadata for a selection of 55 Australian broadcast news videos that need to be downloaded separately. Further information on the SAIVT-BNEWS database is available in our paper.


This paper is also available alongside this document in the file 'Ghaemmaghami2013, Speaker Attribution of Australian Broadcast News Data.pdf'.


The SAIVT-BNEWS ground truth information and associated metadata is licensed CC-BY-SA, and the 55 Australian broadcast news videos (downloaded separately, instructions below) are copyright All Rights Reserved by Fairfax Media.


Attribution


To attribute this database, please include the following citation:
Ghaemmaghami, H., Dean, D., & Sridharan, S (2013) Speaker attribution of Australian broadcast news data, In "Proceedings of the First Workshop on Speech, Language and Audio in Multimedia (SLAM)", CEUR Workshop Proceeedings, Volume 1012, Sun SITE Central Europe, Marseille, France, pp 72-77, available at eprints.

Installation

  1. Download SAIVT-BNEWS.zip
  2. Unzip SAIVT-BNEWS.zip and you should have the following folder structure:
SAIVT-BNEWS
+-- The_Sydney_Morning_Herald_MRSS_Feed
| +--  (for each video)
| +-- *.txt (video metadata)
| +-- *.diarref.lab (speech and speaker
| | ground truth)
| +-- *.faceref.lab (face timing ground truth)
| +-- *.facepositions (face position ground truth)
| +-- *.ocrref.lab (ocr ground truth)
|
+-- code (python script to help download videos)
+-- docs (this file and the publication)

​At this point, you have the SAIVT-BNEWS ground truth information and associated meta data. To download the associated videos, the urls can be found using the information in the *.txt files on the lines starting with 'media_content:', and a python script is provided in the code folder to automate this process. Simply run 'python code/downloadvids.py' to do so.


The videos will be downloaded into the appropriate SAIVT-BNEWS/The_Sydney_Morning_Herald_MRSS_Feed/ folders.


If you aren't using the python script to download the videos, please ensure that only one files is downloaded at a time, and pause briefly between videos to ensure that the media provided doesn't blacklist your IP adress.


Data description


Video metadata


Contains information about the video, including title, a short summary, a link to the video's web page (link), as well as a link to the video itself (media_content), and the id.
This file has one line per each field, with the field name and the value separated by a ':'
-- Example (3123523_high.mp4.txt) --
updated: Wed, 14 March 2012 09:47:50
title: Carr crashes into Senate
summary: After being officially sworn into the Senate, former premier Bob Carr unleashed on the Opposition.
link: http://media.smh.com.au/news/national-news/carr-crashes-into-senate-3123523.html media_content: http://mediadownload2.f2.com.au/flash/media/2012/03/13/3123523/3123523_high.mp4 id: 3123523 ------------------------------------


Diarisation ground truth


Contains information about the speakers appearing on the audio track of the video, as well as a transcription of their speech. 
Each line has the start and end time of the speech (in seconds) followed by a database-level unique speaker identity and finally the speech transcription. 
There may be comments, that should be ignored, indicated by a '#' in the first column, and a commented header to indicate the overall length of the video (in seconds). 
-- Example (3123523.diarref.lab) -- 
#length=100.14 
3.444518 10.693765 paul_bongiorno BACK INTO THE FRAY BOB CARR SWORN IN AS SENATOR SO HE CAN FULFIL A LONG TIME DREAM TO BECOME FOREIGN MINISTER
10.693765 12.922571 bob_carr I WILL BE FAITHFUL A BE A TRUE ALLEGIANCE 
13.035137 17.312643 paul_bongiorno THE BIPARTISAN WELCOME WON'T DETER HIM FROM BEING A GOVERNMENT BOMB THROWER 
17.312643 18.618408 bob_carr TONY ABBOTT IS LIKE THE< br /> --- ... continues ... -------------


Face ground truth


Contains information about the faces appearing in the video. Only faces judged to be sufficiently prominent and frontal are labelled at this stage. 
Each line has the start and end time of the face appearance (in seconds) followed by a database-level unique speaker identity. Identity labels are shared between faces and speakers if they are the same person. 
There may be comments, that should be ignored, indicated by a '#' in the first column, and a commented header to indicate the overall length of the video (in seconds). 
-- Example (3123523.faceref.lab) -- 
#length=100.14 
2.96 6.64 bob_carr 
10.92 14.2 bob_carr 
14.2 15.36 bob_carr 
-- ... continues .. --------------- 
While this file indicates the timing information of the faces in the videos, it does not contain the actual locations of the faces in the video. That information is in the matching *.facepositions file, where each line has the time, the faceid, and the top, left, height and width of the face, collected around 2.5 times per second (or every 10 frames) whenever a face is visible. 
-- Example (3123523.facepositions) -- 
#time id top left height width 
2.96 bob_carr 74 526 76 64 
3.36 bob_carr 100 508 70 60 
3.76 bob_carr 92 526 72 62 
4.16 bob_carr 68 500 70 62 
-- ... continues ... ----------------


OCR Ground truth


Contains information about the on-screen text appearing in the video. Only text appearing in the lower third of the video is considered. 
Each line has the start and end time of the text appearance (in seconds) followed by a video-level unique ocr identity. The identity is designed to indicate when different lines of text appear in the same area within the video. 
There may be comments, that should be ignored, indicated by a '#' in the first column, and a commented header to indicate the overall length of the video (in seconds). 
At this stage the ocr reference does not indicate the location of the ocr text. This may be provided in the future, and/or QUT would be happy to incorporate this information back into the ground truth if it is produced by other researchers. 
-- Example (3123523.ocrref.lab) -- 
#length=100.14 
5.3 7.7 OCR_1 PAUL BONGIORNO
5.3 7.7 OCR_1 NATIONAL AFFAIRS EDITOR 
19.1 21.6 OCR_2 BOB CARR 
19.1 21.6 OCR_2 FOREIGN MINISTER 
-- ... continues .. ---------------


Geographical area of data collection

text
N/A

Publications

Ghaemmaghami, Houman, Dean, David, & Sridharan, Sridha (2013) Speaker attribution of Australian broadcast news data. In Proceedings of the First Workshop on Speech, Language and Audio in Multimedia (SLAM): CEUR Workshop Proceedings, Volume 1012, Sun SITE Central Europe , Marseille, France, pp. 72-77. http://eprints.qut.edu.au/63498/

Research areas

Diarisation; Face Recognition; Speaker Recognition; Clustering; Multi-Modal Recognition; Broadcast News;

Cite this collection

QUT SAIVT: Speech, audio, image and video technologies research . (2016): SAIVT-BNEWS. [Queensland University of Technology]. http://dx.doi.org/10.4225/09/588589af4ae1b

Licence

The SAIVT-BNEWS ground truth information and associated metadata is licensed CC-BY-SA, and the 55 Australian broadcast news videos (downloaded separately, instructions below) are copyright All Rights Reserved by Fairfax Media.

Copyright

©QUT, 2013. The SAIVT-BNEWS ground truth information and associated metadata is licensed CC-BY-SA.

Connections

Has principal investigator
Simon Denman  (Researcher)

Contacts

Name: Dr Simon Denman
Phone: +61731389329

Other

Date record created:
2016-07-01T09:51:48
Date record modified:
2017-01-23T14:43:28
Record status:
Published - Open Access