Always, some individuals in images are more important/attractive than othersin some events such as presentation, basketball game or speech. However, it ischallenging to find important people among all individuals in images directlybased on their spatial or appearance information due to the existence ofdiverse variations of pose, action, appearance of persons and various changesof occasions. We overcome this difficulty by constructing a multipleHyper-Interaction Graph to treat each individual in an image as a node andinferring the most active node referring to interactions estimated by varioustypes of clews. We model pairwise interactions between persons as the edgemessage communicated between nodes, resulting in a bidirectionalpairwise-interaction graph. To enrich the personperson interaction estimation,we further introduce a unidirectional hyper-interaction graph that models theconsensus of interaction between a focal person and any person in a localregion around. Finally, we modify the PageRank algorithm to infer theactiveness of persons on the multiple Hybrid-Interaction Graph (HIG), the unionof the pairwise-interaction and hyperinteraction graphs, and we call ouralgorithm the PersonRank. In order to provide publicable datasets forevaluation, we have contributed a new dataset called Multi-scene ImportantPeople Image Dataset and gathered a NCAA Basketball Image Dataset from sportsgame sequences. We have demonstrated that the proposed PersonRank outperformsrelated methods clearly and substantially.
展开▼