Human Role Assignement in Social Events

We encounter with the problem of role assignment to different persons in video events. Social role can be understood by human interactions and is a key part of human event description. We propose supervise framework here, where role labels are available in training phase. To model problem Conditional Random Field is used because inter-role interaction of different characters is key part of problem. Some features extracted from appearance, movement and different aspects of person which is called unary features. Also some features are extracted from inter-role interaction of people which is called pairwise features. Unary and pairwise features are needed to train CRF model. A new marginal based learning method is used for model parameter estimation. For model evaluation, social role dataset which is collected from YouTube website is used. Performance of model is compared with some baselines including Ramanathan CRF Model which is a state of the art model in this area.

Keywords:

component: Social Role , Supervise Framework , CRF , unary and pairwise features