Social interaction detection by computer models in movement trajectories

Abstract

Even when watching animations of two dots moving on a white background, people often easily and automatically interpret what they see in terms of social interactions. Although the mechanisms by which people detect social interactions are not fully understood yet, recently, different cognitive computer models have been developed that fit human data well. The research question of this study is whether a bottom-up, hierarchical model designed by Shu et al. (2018) based on subinteraction classification and temporal interactivity parsing can explain human social interaction from movement trajectories. We designed a simple computer game which we used to generate stimuli trajectories and animations. In a second experiment participants rated the perceived social interaction in these animations. In line with our expectations, the hierarchical model fitted these experimental data well. However, average proximity between objects served as an almost equally good indicator for social interaction and an exponential model based on average proximity performed even slightly better. Using the principle of Occam’s razor, we conclude that in videos of moving shapes, a simple proximity based approach offers a better explanation for human social interaction detection than the proposed hierarchical model.