With an explosion in multimedia content production due to ever increasing reach of internet, content in the form of videos is becoming increasingly commonplace, becoming the preferred method for the purpose of delivering content. Advent of social media and growth of video sharing websites, in particular Youtube, has only contributed to the increasing importance of videographic content. More content is uploaded to Youtube a day, than a person is capable of watching in his/her whole lifetime. With the emergence of video content as an effective mode of information propagation, automating the process of summarization a video has become paramount. Video Summarization, in recent times, has emerged as a challenging problem in the field of machine learning, which aims at automatically evaluating the content of a video, and generating a summary with the most relevant content of the video. Video summarization finds applications in generating highlights for sports events, trailers for movies and in general shortening video to the most relevant subsequences, allowing humans to browse large repository of videos efficiently.
Video summarization is a challenging problem in multiple facets. There is no natural ordering of video summaries. Among a given set of video summaries, the best representation of the original video is highly subjective. The general objective in modern literature for video summarization aims at producing a summary, typically 5%-15% of the whole video, consisting of the most informative content from the original video. The content is usually represented in the form of ”key frames”, or more appropriately as video skims. A good video summary depicts the synopsis of the original video, in a compact way depicting all important and relevant scenes/shots. Through this project, we review the major techniques in video summarization, and look at their performance on some recent datasets. Most of the work pertaining to this project can be found on github here