Survey and Comparison of Video Summarization Techniques

Abstract

Video summarization is a keenly intellective video compression technique to select a subset of keyframes or keyshot which are combined to represent shorter and compendious summary of the original input video without losing the contextual semantics of the same. In context of summary generated, it can be divided into static (static storyboard) and dynamic (video skimming) video summarization. Supervised ,unsupervised and reinforcement learning techniques have been introduced in the literature of video summarization. Earlier development in this field marked the introduction of various unsupervised techniques that used handcrafted heuristics to select mutually independent keyframes. In recent years, supervised techniques and deep reinforcement learning techniques have been introduced that models the structural semantics of the original video to generate summaries using frame - level ground truth generated by humans, so that we get the result as close as possible to the human understanding of the video. This paper aims to introduce different architectures proposed for video summarization and provides a qualitative and quantitative comparison of these methods.