Hierarchical HOI Detection Framework Augmented by Human Interactive Intention

Human-Object Interaction (HOI) detection is a core task for human-centric visual relationship understanding. However most existing methods still face the challenge of accuracy improvement. Because there are usually too many human-object pairs and complex scenes in the image, which leads to the complexity of HOI detection task. In this paper, we innovatively introduce a multi-grained interactive intention layer to deal with human gaze and multi-level human pose information in complex human-object scenes. We construct a hierarchical architecture using macro-level spatial layout cues, mesco-level interactive intentions and micro-level appearance features of humans and objects to complement the various levels of information. Leading results on two public benchmarks validate the effectiveness of our model.