CMU-SMU@TRECVID 2015: Video Hyperlinking
Zhiyong Cheng, Xuanchong Li, Jialie Shen, Alexander G. Hauptmann · 2015
In this report, we describe CMU-SMU’s participation in the Video Hyperlinking task of TRECVID 2015. We treat video hyperlinking as ad-hoc retrieval scenario and use a variety of retrieval methods. Our experiments mainly focus on the study of different features on the performance of video hyperlinking, including subtitle, metadata, audio and visual features, as well as the consideration of surrounding context. Different combination strategies are used to combine those features. Besides, we also attempt to categorize the queries and use different search strategies for different categories. Experiments results show that (1) the context does not generally improve results, (2) the search performance mainly rely on textual features, and the combination of audio and visual feature cannot provide improvements; (3) due to the lack of training examples, machine learning techniques cannot provide contributions.