@INPROCEEDINGS{1324Kelm2011,
	AUTHOR = {Pascal Kelm and Sebastian Schmiedeke and Kai Clüver and Thomas Sikora},
	TITLE = {Automatic Geo-referencing of Flickr Videos},
	BOOKTITLE = {NEM Summit 2011},
	YEAR = {2011},
	MONTH = sep,
	EDITOR = {Eurescom – the European Institute for Research and Strategic Studies in Telecommunications   – GmbH},
	PUBLISHER = {Sigma Orionis},
	PAGES = {76--80},
	ORGANIZATION = {NEM  Summit},
	ADDRESS = {Torino,  Italy,  September  27-29  2011},
	NOTE = {Copyright © 2011 – Eurescom GmbH – On behalf of NEM Initiative
All  rights  on  Proceedings  of  2011  NEM  Summit  (Torino,  Italy,  September  27-29  2011)  
reserved.  All  rights  on  individual  papers,  published  in  the  proceedings,  remain  unaffected.

ISBN 978-3-00-035465-6


Publisher
Eurescom – the European Institute for Research and Strategic Studies in Telecommunications 
– GmbH
Wieblinger Weg 19/4 - 69123 Heidelberg - Germany
Phone: +49 6221 989 0 - Fax: +49 6221 989 209 - http://www.eurescom.eu
For publisher: Halid Hrasnica
On behalf of NEM Initiative – http://www.nem-initiative.org 


eBook and USB produced by Sigma Orionis
1240, route des dolines - BP287 Valbonne - France
Phone: +33 (0) 493 001 550 - Fax: +33 (0) 493 001 560 - http://www.sigma-orionis.com
For producer: Florent Genoux
On behalf of NEM Initiative – http://www.nem-initiative.org},
	PDF = {http://elvera.nue.tu-berlin.de/files/1324Kelm2011.pdf},
	ABSTRACT = {We present a hierarchical, multi-modal approach for geo-referencing Flickr videos. Our approach makes use of external resources to identify toponyms in the metadata and of visual features to identify similar content. We use a database of more than 3.6 million Flickr images to group them into geographical areas and to build a hierarchical model. First, the geographical boundaries extraction method identifies the country and its dimension. Then, a visual method is used to classify the videos’ location into plausible areas. Next, the visually nearest neighbour method is used to find correspondences with the training images within the pre-classified areas. As the processed video sequences are represented using low-level feature vectors from multiple key frames, we also present techniques for video to image matchings. The Flickr videos are tagged with the geo-information of the visually most similar training item within the areas previously filtered in the pre-classification step. The results show that we are able to tag one third of our videos correctly within an error margin of 1 km.}
}