Television Advertisement Analysis Using Attention-based Multimodal Network

Abstract

The impression/emotion (e.g. the recognition rate, favorableness) prediction of an advertisement is important. It is related to multimodal features including frames, sounds, as well as metadata. In this paper, we propose a system that can utilize different models to embed different features, and apply attention mechanism to efficiently combine those features to help predict the impressions/emotions of audience after they watch an advertisement. Our prediction can achieve the state-of-the-art performance in real-world dataset. This system can also detailed analyze the importance of advertisement components.