Dynamic Clustering Algorithm for Video Summarization on VSUMM Dataset

Abstract

The extent of videos being produced per day across the world is enormous, and in the past few years it has increased to an unprecedented level. Information extraction from a video, however, is more difficult than information extraction from an image. A viewer must see the entire video in order to understand its context. Aside from context awareness, it is nearly impossible to make a universally applicable summary video because each person has a different preferred keyframe. A number of approaches came into existence for tackling this problem which include supervised and unsupervised learning techniques, and some associated with Deep Learning techniques. However, it would require a significant amount of individualized data labelling if we attempted to approach problem video summarizing via a supervised learning method. In this paper, we developed an algorithm based on Dynamic Clustering of projected frame histograms approach to address the challenge of video summarization using unsupervised learning. We have tested the performance of the approach on the VSUMM, a benchmark dataset and showcased that using dynamic clustering algorithm has been proven to perform competitively with some existing approaches.