Complete interest propagation from part for visual relation of interest detection

You Zhou, Fan Yu

Abstract: Visual relation detection (VRD) is proposed to describe an image with visual relation triplets in the form of <subject, predicate, object>. As a further extension of the traditional VRD task, visual relation of interest detection (VROID) is proposed to obtain visual relations of interest, i.e., visual relations are semantically important for expressing the main content of an image. In this paper, we propose a complete interest propagation from part (CIPFP) method for VROID, which exploits semantic parts and propagates interest along part-instance-relation. Specifically, the interest in CIPFP is propagated from parts to part pairs, from parts to instances, from part pairs to instance pairs, from instances to instance pairs, from parts to relation triplets and from instance pairs to relation triplets. We conduct substantial experiments to validate the effectiveness of the CIPFP method and the components in CIPFP.