An Improved Dominant-Set Algorithm for Detecting Conversational Groups in Images

Abstract

Social encounters modelling is gaining more prominent as the need to analysis diverse social interaction groups is on the rise in both videos and images. Inspired by social psychological, social encounter attempts to analysis who is interacting with whom in a social gathering such as party, chat, etc. There exist two types of social encounters namely: focused and unfocused encounters. In this thesis, we modelled social encounter using a game-theoretical framework. We translate the notion of evolutionary game theory into social encounter where the pure strategy corresponds to the detected individuals in the scene and the payoff corresponds to similarity measure between subjects. Using this approach with position and orientation we are able to statistically modelled F-formation. We experimented this approached on several benchmark datasets and our experiment shows significant improvement in performance.