TY - GEN
T1 - Exploiting visual-based intent classification for diverse social image retrieval
AU - Wang, Bo
AU - Larson, Martha
PY - 2017
Y1 - 2017
N2 - In the 2017 MediaEval Retrieving Diverse Social Images task, we (TUD-MMC team) propose a novel method, namely an intent-based approach, for social image search result diversification. The underlying assumption is that the visual appearance of social images is impacted by the underlying photographic act, i.e., why the images were taken. Better understanding the rationale behind the photographic act could potentially benefit social image search result diversification. To investigate this idea, we employ a manual content analysis approach to create a taxonomy of intent classes. Our experiments show that a CNN-based neural network classifier is able to capture the visual difference between the classes in the intent taxonomy. We cluster images of the Flickr baseline based on predicted intent class and generate a re-ranked list by alternating images from different clusters. Our results reveal that, compared to conventional diversification strategies, intent-based search result diversification is able to bring a considerable improvement in terms of cluster recall with several extra benefits.
AB - In the 2017 MediaEval Retrieving Diverse Social Images task, we (TUD-MMC team) propose a novel method, namely an intent-based approach, for social image search result diversification. The underlying assumption is that the visual appearance of social images is impacted by the underlying photographic act, i.e., why the images were taken. Better understanding the rationale behind the photographic act could potentially benefit social image search result diversification. To investigate this idea, we employ a manual content analysis approach to create a taxonomy of intent classes. Our experiments show that a CNN-based neural network classifier is able to capture the visual difference between the classes in the intent taxonomy. We cluster images of the Flickr baseline based on predicted intent class and generate a re-ranked list by alternating images from different clusters. Our results reveal that, compared to conventional diversification strategies, intent-based search result diversification is able to bring a considerable improvement in terms of cluster recall with several extra benefits.
UR - http://www.scopus.com/inward/record.url?scp=85035074679&partnerID=8YFLogxK
UR - http://resolver.tudelft.nl/uuid:c15a255b-65b6-4328-8df7-9f579c48b14d
M3 - Conference contribution
AN - SCOPUS:85035074679
T3 - CEUR Workshop Proceedings
SP - 1
EP - 3
BT - Working Notes Proceedings of the MediaEval 2017 Workshop
A2 - Gravier, Guillaume
A2 - Bischke , Benjamin
A2 - Demarty, Claire-Hélène
A2 - Zaharieva, Maia
A2 - Riegler, Michael
A2 - Dellandrea, Emmanuel
A2 - Bogdanov, Dmitry
A2 - Sutcliffe, Richard
A2 - Jones, Gareth J.F.
A2 - Larson, Martha
T2 - MediaEval 2017
Y2 - 13 September 2017 through 15 September 2017
ER -