Sitcom-stars oriented video advertising via clothing retrieval

Abstract

This paper introduces a novel learning-based framework for video content-based advertising, DeepLink, which aims at linking sitcom-stars and online stores with clothing retrieval by using state-of-the-art deep convolutional neural networks (CNNs). Concretely, several deep CNN models are adopted for composing multiple sub-modules in DeepLink, including human-body detection, human-pose selection, face verification, clothing detection and retrieval from advertisements (ads) pool that is constructed by clothing images collected from real-world online stores. For clothing detection and retrieval from ad images, we firstly transfer the state-of-the-art deep CNN models to our data domain, and then train corresponding models based on our constructed large-scale clothing datasets. Extensive experimental results demonstrate the feasibility and efficacy of our proposed clothing-based video-advertising system.