Achieving real-time video summarization on commodity hardware

Abstract

We present a system for automatic video summarization which is able to operate in real-time on commodity hardware. This is achieved by performing segmentation to divide a video into a series of small video clips, which are further reduced or eliminated with the assistance of highly efficient low-level features. A numerical score is then assigned to each segment by our model trained using a set of highperformance hand-crafted features. Finally, segments are selected based on their score to generate a final video summary. On our benchmark dataset, we achieve results competitive to other methods. In cases where our accuracy is lower than competitive methods, we achieve significantly higher performance. We additionally present methods for generating additional summaries almost instantly, and for learning user preferences over time—two processes which are often overlooked in work on video summarization, but essential for real-world use