Selecting an Iconic Pose From an Action Video

Abstract

This paper presents a method for selecting an iconic pose frame from an action video. An iconic pose frame is a frame showing a representative pose, distinct from other actions. We first extract a diverse set of keyframes from the video using unsupervised video summarization. A classification loss ensures that the selected frames retain high action classification accuracy. To find iconic poses, we introduce two loss terms, an Extreme Pose Loss, encouraging selecting poses far from the mean pose, and a Frame Contrastive Loss, which encourages poses from the same action to be similar. In a user preference study on UCF-101 videos we show that the automatically selected iconic pose keyframes are preferred to manually selected ones in 48% of cases.