Dynamic and Personalized Video Summarization Based on User Preferences

Personalized video summarization entails generation of short and compact video summary based exclusively on user preferences. The perception of individual user preferences and selection of relevant and salient video content requires domain knowledge understanding. This paper presents a stacked Convolutional Neural Network (CNN) approach by embedding action-based and rule-based domain knowledge of Cricket sport. The proposed approach relies on a stacked organization of Umpire detection and Umpire pose recognition modules, followed by domain rules to select video content matching the user preferences. This content selection strategy facilitates the generation of dynamic and personalized video summary grounded on domain knowledge. This paper also proposes two novel summary evaluation metrics based on user reactions, i.e., User Rating Score and Composite Summary Score (CS-Score). The proposed approach performs exceptionally well with deep action-based features and rule-based features for better perception and understanding of user preferences and synchronization-selection of relevant video content. The results indicate promising performance of both models and present a standard and benchmark platform for sports, personalized and dynamic video summarization.