Labelling Human-Object-Interaction images by attribute based modelling

In this research, we have proposed a novel attribute-centric approach towards classifying Human-Object-Interactions (HOI) from RGBD images. Unlike the other contemporary methods in HOI classification, our approach takes into consideration the physical attributes of the Human, Object, and Ambience (nearby object) contexts to classify HOI images robustly. We argue that the attributes (i.e. object appearance, object shape, object material, human body poses, etc.) are distributed and shared between different HOI classes and thus give a compact and discriminative representation of the HOI.