Classifying advertising video by topicalizing high-level semantic concepts

摘要

The recent proliferation of videos has driven the research into various applications, ranging from video analysis to indexing and retrieval. These applications greatly benefit from domain knowledge of videos. As a special kind of videos, classifying ad video is a key task because it allows automatic organization of videos according to categories or genres, and this further enables ad video indexing and retrieval. However, classifying ad video is challenging due to its unconstraint content and distinctive expression. While many studies focus on selecting ads relevant to the target videos, to the best of our knowledge, few focuses on ad video classification. To classify ad video, we propose a novel video representation that aims to capture the latent semantics of ad video in an unsupervised manner. In particular, this paper integrates the posterior occurrence probability between brand/logo information and the high-level object information into a latent Dirichlet allocation unified learning paradigm, named ppLDA. A topical representation for ad video is obtained by the proposed method, which can support category-related task. Our experiments on 10,111 real-world ad videos downloaded from Internet demonstrate that the proposed method could effectively differentiate ad videos.