FOREGROUND-CENTRIC ACTION RECOGNITION

Abstract

Video is highly unstructured and high dimensional data. While action recognition has been an active research topic, modeling structures in video is often a challenging problem. Modeling structures would reduce search spaces of learning problems in high dimensional video data and produce competitive performances. Inspired by these prospects, in this thesis, we attempt to develop structural action models based on a robust and prominent cue in video, figure-ground separation. Going through a paradigm shift in computer vision community, we also develop structural models powered by both shallow and deep predictive machinery.