Efficient Web Video Classification via Cross-modality Knowledge Transferring

This paper puts forward a novel method for classifying Web videos with high efficiency. Instead of analyzing the videos or extracting complicated visual features, which are both computationally expensive, we only utilize the related textual information of the to-be-classified Web videos. To address the sparsity of the textual features, we propose to exploit knowledge from auxiliary data of diverse modalilies during training, such that more informative features can be constructed. We carried out extensive experiments on MCG-WEB dataset collected from YouTube for video classification. The results demonstrate that our method can outperform several related state-of-the-art methods markedly and is quite fast, validating its effectiveness and efficiency.