The France Telecom Orange Labs (Beijing) Video Semantic Indexing Systems - TRECVID 2012 Notebook Paper.
Kun Tao, Yuan Dong, Yunlong Bian, Xiaofu Chang, Hongliang Bai, Wei Liu, Feng Zhao, Peng Li, Chengbin Zeng · 2012
In this paper, we describe FTRDBJ’s systems and experiments for TRECVID2012 SIN task. This year, a new type of runs named “concept pair ” was evaluated, for which we tried many different models. We submitted two “full ” runs and two “pair ” runs. For the “full ” runs, our systems are very similar to the old ones used in 2011. A 6-feature and a 9-feature composite-kernel SVM systems were used to detect the concepts. For the “pair ” runs, we submitted a confidence based run and an OWA based run. The final results show that the first one worked better because a simpler model usually works more robust. 1.