Recognising Human-Object Interactions Using Attention-based LSTMs

Abstract

Recognising Human-object interactions (HOIs) in videos is a challenge task especially when a human can interact with multiple objects. This paper attempts to solve the problem of HOIs by proposing a hierarchical framework that analyzes human-object interactions from a video sequence. The framework consists of LSTMs that firstly capture both human motion and temporal object information independently, followed by fusing these information through a bilinear layer to aggregate human-object features, which are then fed to a global deep LSTM to learn high-level information of HOIs. The proposed approach applies an attention mechanism to LSTMs in order to focus on important parts of human and object temporal information.