Video distillation
Xuelong Li, Bin Zhao · Scientia Sinica Informationis · 2021
Video has become one of the most important data forms. Video distillation explores more compact data forms and information modalities by analyzing the spatial-temporal and semantic features of video data, which is an important task in computer vision and a key technique in artificial intelligence. With the rapid development of video capturing devices and the increasing human requirements, video analysis tasks are facing numbers of opportunities and challenges. In recent years, large amounts of video distillation approaches are proposed. This paper creatively unifies the theoretical basis of video distillation by analyzing the relationship among data, information and knowledge from the perspective of information theory, and argues that the principle of video distillation is to improve the information capacity of video data. Then, we overview existing approaches in the aspects of video data representation, key content summarization, moving object synopsis and text description generation, etc., and relate the development of video summarization, synopsis and captioning, which are typical tasks in video distillation. More importantly, this paper discusses the advantages and drawbacks of existing approaches, and then points out several key scientific problems that have not yet been addressed, and simultaneously analyzes the potential future development in video distillation.