An Improved Human-Object Interaction Detection Network

Human-Object Interaction (HOI) Detection is an important problem which is expected not only to detect individual object instances, but also to recognize the visual relationship between object pairs. In this paper, we improve the performance of the HOI detection task based on the instance-centric attention network. At first, we improve the accuracy of existing model by optimizing loss function and training details. We validate the improved method on the recently introduced Verbs in COCO (V-COCO) dataset. Then, to analyze the experimental results, we set several error modes and get the distribution of the false positives for each action class. Finally, we apply the trained model to HOI detection in surveillance scenario. The experimental result shows that the improved method can accurately detect the human-object interactions, which are involved in the VCOCO dataset