Live Video Action Recognition from Unsupervised Action Proposals

The problem of action detection in untrimmed videos consists in localizing those parts of a certain video that can contain an action. Typically, state-of-the-art approaches to this problem use a temporal action proposals (TAPs) generator followed by an action classifier module. Moreover, TAPs solutions are learned from a supervised setting, and need the entire video to be processed to produce effective proposals. These properties become a limitation for certain real applications in which a system requires to know the content of the video in an online fashion. To do so, in this work we introduce a live video action detection application which integrates the action classifier step with an unsupervised and online TAPs generator. We evaluate, for the first time, the precision of this novel pipeline for the problem of action detection in untrimmed videos. We offer a thorough experimental evaluation in ActivityNet dataset, where our unsupervised model can compete with the state-of-the-art supervised solutions.