The SUM-GAN-sl method for video summarization

Say hello to SUM-GAN-sl, a deep learning architecture for unsupervised video summarization, that doesn’t require the existence of extensively annotated data.

Video summarization provides a short visual summary that encapsulates the flow of a story and the essential parts of an original full-length video. Automating the creation of video summaries has many uses, e.g. for media organizations and online video archives to allow effective indexing, browsing, retrieval and promotion of their media assets. It also enables you to post summaries on a variety of video sharing platforms. This improves the viewing experience, enhances viewer engagement and increases content consumption. Ultimately, being able to easily create multiple summaries for a single video means that you can share your content across channels and devices, in versions that are tailored to the needs of each audience and channel.

Our method is capable of learning how to identify the most representative parts of a video without the use of ground-truth summaries. It can be easily adapted to different types of video content by being trained on an adequately large and representative set of data.

An overview of the first architecture that we developed can be found in our AI4TV@ACMMM2019 presentation. Full details available in our AI4TV paper, at the publisher’s website and Zenodo.