NBA Basketball Video Summarization for News Report via Hierarchical-Grained Deep Reinforcement Learning.

At present, the demand for short video generation is increasing, especially for sports news report, which urgently needs automatic video summarization methods to reduce time and labor cost. This paper focuses on NBA basketball videos and seeks for the actual needs of news report on sports video summarization. We propose a hierarchical-grained deep reinforcement learning framework to generate short basketball video. For a long basketball game video, we propose a hierarchical-grained subshot segmentation algorithm, which takes into account both semantics and objective factors, and preserves spatiotemporal consistency. Then we select candidate frames through a news element enhanced deep reinforcement learning framework. On this basis, a news report oriented video summarization algorithm based on probability sampling is implemented with the fusion of multi-game and multi-news elements. Experimental results on the NBA dataset newly collected by us demonstrate the effectiveness of the proposed framework. Moreover, the proposed method is able to highlight the video content including well preserved news elements.