Human-Object Interaction Detection with Missing Objects

Existing studies on human-object interaction (HOI) assume that human and object instances can be detected. This paper proposes a more practical HOI detection method for when object instances are not necessarily easily detectable. To our knowledge, we introduce the first method for such challenging HOI detection that incorporates global scene information. The two most widely used public HOI benchmark datasets are shown to contain many cases of HOI with missing objects (HOI-MO). We label these to compose new test sets for the proposed method. The effectiveness and superiority of the proposed method are demonstrated through extensive experiments and comparisons.