Context-guided Unsupervised Video Summarization Network

Video summarization aims to generate concise static video frame collection or dynamic video browsing. Regarding this, retaining key frames as much as possible and removing redundant ones is a key issue. To this end, this paper develops an unsupervised context-guided video network (CGVN). In specific, the long-distance inter-frame reliance relationship can be established via the multi-view self-attention mechanism, which solves the problem of memory limitations of traditional methods on the process of long sequence. At the same time, the unique parallel processing method has also greatly reduced the cost of training. Moreover, the video summarization is taken as an unsupervised process of continuous decision-making under the reinforcement learning framework. Based on the joint diversity and representativeness loss, the summarization can be learned, which avoids excessive dependence on user labeling, so that the subjectivity of summarization is too strong and lacks objective attributes. Comparative experiments on two public benchmarks, i.e., SumMe and TVSum are conducted. The experimental results demonstrate the effectiveness of our proposed method.