Complex event detection via attention-based video representation and classification

摘要

As an important task in managing unconstrained web videos, multimedia event detection (MED) has attracted wide attention recently. However, due to the complexities such as high abstraction of the events, various scenes and frequent interactions of individuals etc., MED is quite challenging. In this paper, we propose a novel MED algorithm via attention-based video representation and classification. Firstly, inspired by human's selective attention mechanism, an attention-based saliency localization network (ASLN) is constructed to quickly predict the semantic saliency objects of video frames. Afterwards, in order to complementarily represent salient objects and the surroundings, two Convolutional Neural Networks (CNNs) features, i.e., local saliency feature and global feature are respectively extracted from the salient objects and the whole feature map. Thirdly, after binding two features together, Vector of Locally Aggregated Descriptors (VLAD) is applied to encode them into the video representation. Finally, the linear Support Vector Machine (SVM) classifiers are trained to classify. We extensively evaluate the performance on TRECVID MED14_10Ex, MED14_100Ex and Columbia Consume Video (CCV) datasets. Experimental results show that the proposed single model outperforms state-of-the-art approaches on all three real-world video datasets, and demonstrate the effectiveness.