Person In Context Challenge

Recent years have witnessed the great progress of the perception task such as image classification, object detection and pixel-wise semantic/instance segmentation. It is the right time to go one step further to infer the relations between the objects. Increasingly more efforts are devoted to relation prediction, such as the Visual Genome and Google Open Image challenge. There are mainly two differences between existing relation prediction works and PIC challenge. Firstly, instead of inferring all the relations between any two objects, PIC focuses on estimating human-centric relations, including human-object relations and human-human relations. Each relation is represented by a triplet in the form of <Subject, Relation, Object>, such as <Human A, hold, Bottle A> and <Human A, hug, Human B>. In other words, under the definition of PIC, ‘Subject’ in the triplet should be human. Secondly, PIC targets at the relation segmentation. More precisely, traditional relation prediction only estimates the bounding box of ‘Subject’ and ‘Object’ while PIC needs to estimate their masks (shapes).