MediaMill at TRECVID 2014: Searching Concepts, Objects, Instances and Events in Video
Cees G. M. Snoek, Koen E. A. van de Sande, Daniel Fontijne, Spencer Cappallo, Jan C. van Gemert, Amirhossein Habibian, Thomas Mensink, Pascal Mettes, Ran Tao, D.C. Koelma, Arnold W. M. Smeulders · UvA-DARE (University of Amsterdam) · 2014
In this paper we summarize our TRECVID 2013 [15] video retrieval experiments. The MediaMill team participated in four tasks: concept detection, object localization, in-stance search, and event recognition. For all tasks the starting point is our top-performing bag-of-words system of TRECVID 2008-2012, which uses color SIFT descrip-tors, average and difference coded into codebooks with spa-tial pyramids and kernel-based machine learning. New this year are concept detection with deep learning, concept detec-tion without annotations, object localization using selective search, instance search by reranking, and event recognition based on concept vocabularies. Our experiments focus on es-tablishing the video retrieval value of the innovations. The 2013 edition of the TRECVID benchmark has again been a fruitful participation for the MediaMill team, resulting in the best result for concept detection, concept detection with-out annotation, object localization, concept pair detection, and visual event recognition with few examples.