Weakly Supervised Video Summarization Method Using Video Popularity

With the increasing popularity and decreasing cost of smart phones and video editing tools,the amount of video data has increased drastically in the past few years.Video has become one of the most important form of visual data.According to You Tube statistics,more than 300 hours of video are uploaded to Youtube every minute.Due to the sheer amount of video data,it is unrealistic for users to watch all these videos and identify useful information.Therefore,developing automatic video summarization to help users efficiently browse videos becomes more and more important.The goal of video summarization is to create a shorter video that captures the important information of the input video.The problem is challenging due to its subjectiveness.Most existing works summarize videos by either designing heuristic criteria in an unsupervised way,or developing supervised algorithms by leveraging human-crafted training data.However,unsupervised methods are blind to the information of the video itself and often fail to produce semantically meaningful video summaries.One the other hand,acquisition of large amount of the human-crafted training data for supervised methods is very hard and may lead to a biased model.To tackle the problems mentioned above,in this paper,we propose a weakly supervised approach that exploits only easy-to-obtain videl-level popularity information(e.g.,clicks,play counts,favorite counts)for summarizing videos,named Popularity-based Deep Summarization Network(PDSN).Specifically,PDSN contains two modules.The first is the summarisation module,we train a summarisation network with policy gradient algorithm.The second is the evaluation module,We train a popularity evaluation network to evaluate the quality of the video summaries to provide rewards for training the summarisation network.Extensive experiments on three benchmark datasets(Sum Me,TVSum and Co Sum)show that the proposed approach achieves state-of-the-art to several recently proposed approaches.The main innovations of this paper are as follows:(1)To the best of our knowledge,we are the first to apply video popularity informations to video summarization.(2)Our proposal is weakly supervised and requires only video-level annotations,which can drastically reduce the cost on annotation.(3)Different to other video summarization methods based on reinforcement learning,our proposal uses the popularity evaluation network as a reward function.The purpose of training is to maximize the expected reward over a period of time by producing high-quality video summaries.(4)Our proposal not only outperforms state-of-the-art unsupervised video summarization methods and most weakly supervised methods,but also some supervised methods.(5)Designing and implementing a video summarization service based on ctr data of the advertising videos,which can maximize the ctr of the advertising video on online advertising platform.This paper consists of six chapters.The first chapter introduces the research status of video summarization and the problems to be solved.The second chapter introduces the specific definitions and related algorithms of video summarization.The third chapter introduces the weakly supervised video summarization method using video popularity.The fourth chapter introduces the detailed experimental results and analysis.The fifth chapter introduces a video summarization service based on ctr data of advertising videos.Finally,the last chapter are the summary and further prospects.