Ventral & Dorsal Stream Theory based Zero-Shot Action Recognition

Abstract

Most Zero-Shot Action Recognition (ZSAR) methods establish visual-semantic joint embedding space, which is based on commonly used visual features and semantic embeddings, to learn the correlation between actions. Nevertheless, extracting visual features without structural guidance would lead to sparse video features, which reflect the correlation of actions, fall into oblivion. Based on the Ventral & Dorsal Stream Theory (VD), we propose a VD-ZSAR method to extract irredundant visual feature, which can relieve relation ambiguity caused by redundant visual feature. And a visual-semantic joint embedding space is learned by combining nonredundant visual space with semantic space. Specifically, visual space is constructed by the motion cues perceived by Dorsal Stream, and the object cues perceived by Ventral Stream. Semantic space is constructed by sentence-to-vector generator. The visual-semantic joint embedding space is built by a nonlinear similarity metric learning mechanism, which can better implicitly reflect the correlation between actions. Extensive experiments on the Olympic, HDMB51 and UCF101 datasets validate the favorable performance of our proposed approach.