Large Scale Real-Life Action Recognition Using Conditional Random Fields with Stochastic Training

Xu Sun, Hisashi Kashima, Ryota Tomioka, Naonori Ueda

Abstract: Action recognition is usually studied with limited lab settings and a small data set. Traditional lab settings assume that the start and the end of each action are known. However, this is not true for the real-life activity recognition, where different actions are present in a continuous temporal sequence, with their boundaries unknown to the recognizer. Also, unlike previous attempts, our study is based on a large-scale data set collected from real world activities. The novelty of this paper is twofold: (1) Large-scale non-boundary action recognition; (2) The first application of the averaged stochastic gradient training with feedback (ASF) to conditional random fields. We find the ASF training method outperforms a variety of traditional training methods in this task.