Comparison and Evaluation of Video Retrieval Approaches Using Query Sentences

Abstract

Following are two mainstream approaches of video retrieval from large-scale video data using query sentences: (1) an approach to find pre-trained concepts such as objects, persons, scenes, and activities corresponding to a query sentence, and (2) an approach to map a query sentence and images/videos into the same feature space and directly search for images/videos that match the query sentence. In this study, we analyze the advantages and disadvantages of these two approaches using a large-scale video database of TRECVID benchmark and confirm whether the fusion of these approaches can improve video retrieval performance.