Abstract : This paper presents the runs that we submitted in the context of the TRECVid 2015 Video Hyperlinking task. The task aims at proposing a set of video segments, called targets, to complement a query video segment defined as anchor. We used automatic transcripts and automatically extracted visual concept as input data. Two out of four runs use cross-modal LDA as a means to jointly make use of visual and audio information in the videos. As a contrast, one is based solely on visual information, and a combination of the cross-modal and visual runs is considered. After presenting the approaches, we discuss the performance obtained by the respective runs, as well as some of the limitations of the evaluation process.
https://hal.archives-ouvertes.fr/hal-01403726 Contributor : Pascale SébillotConnect in order to contact the contributor Submitted on : Sunday, November 27, 2016 - 2:11:49 PM Last modification on : Wednesday, February 2, 2022 - 3:50:39 PM Long-term archiving on: : Tuesday, March 21, 2017 - 10:06:22 AM
Anca-Roxana Simon, Ronan Sicre, Rémi Bois, Guillaume Gravier, Pascale Sébillot. IRISA at TrecVid2015: Leveraging Multimodal LDA for Video Hyperlinking. TRECVid 2015 Workshop, Nov 2015, Gaithersburg, United States. ⟨hal-01403726⟩