PROJECT TITLE :

Transferable Interactiveness Knowledge for Human-Object Interaction Detection

ABSTRACT:

In order to gain a better understanding of the ways in which people interact with things around them, it is necessary to solve the problem of human-object interaction (HOI) detection. In this paper, we investigate the concept of interactiveness knowledge, which describes the ability of human and non-human objects to interact with one another. We came to the conclusion that interactiveness knowledge could be learned across multiple HOI datasets, which would help reduce the disparity that exists between the various HOI category settings. Our primary goal is to learn the general interactiveness knowledge from a variety of HOI datasets using an Interactiveness Network and then perform Non-Interaction Suppression prior to HOI classification using inference. Because of the generalization of interactiveness, the interactiveness network is a knowledge learner that can be transferred to other contexts and can work in conjunction with any HOI detection model to achieve the desired outcomes. In order to learn the interactiveness in hierarchical paradigm, that is, instance-level and body part-level interactivenesses, we make use of both the human instance and body part features in conjunction with one another. Following this, a consistency task is suggested as a means of guiding the learning process and eliciting deeper interactive visual clues. We conduct in-depth evaluations of the proposed method on three different datasets: HICO-DET, V-COCO, and a freshly created HAKE-HOI dataset. Our method outperforms other state-of-the-art HOI detection methods, which demonstrates both its efficiency and its adaptability. This is due to the interactiveness that we have learned.