Deep Learning for Video Localization

The great success of video recognition has contributed to many real-world applications such as automatic video tagging for video websites and apps. However, video recognition is limited in understanding the overall event that exists in a video, without a fine-grained analysis of video segments. To compensate for the limitations of video recognition, video localization provides an accurate and comprehensive understanding of videos by predicting when and where an action occurs in a video. This chapter will introduce two representative tasks of video localization including action localization and temporal video grounding. Action localization aims to find the video segments that contain potential actions and predict the action classes, while temporal video grounding aims to localize video moments that best match given natural language. We present an overview of existing approaches and benchmarks used for evalution.