AI-Cafe presents: Deep Learning for Video Summarization

Nowadays, both social media sites, Media Archiving and Management (MAM) systems, and broadcaster mediatheques store huge amounts of video information. Its search can be very tedious, whether for journalistic purposes or just for selecting a video to watch and entertain. Therefore, fast video browsing is an essential service to be provided in a www environment. One way to do it is through video summarization. It can create a picture gallery of each video, thus greatly facilitating its browsing. This gallery should contain both representative and informative key video frames. The same can be done by video skimming, which retains only the most representative and informative video snippets. Movie trailers is a primary example of video skimming. Recently, deep learning (DNN) techniques have been extensively employed in video summarization. They typically exploit both spatial and temporal video information, e.g., by using combined CNN and LSTM video modeling approaches. Adversarial DNN and GAN methods have also been used in video summarization. This lecture will overview all recent advances in the use of Deep learning for video summarization.