Abstract
Efficient and accurate judgments about threat from dyadic interactions are important for survival in the social world. Systematically assessing the separate impact of body movements and context has proven difficult due to the graphic nature of threatening displays and level of emotion involved. Here, we employed advanced computer vision algorithms to alter the videos, and assess the contribution of kinematic information and contextual information to support recognition of threatening actions. Experiment 1 presented 30 YouTube videos depicting a range of threatening and non-threatening interactions. Participants were asked to perform threat classification and provide written descriptions. In Experiment 2, twenty-six of the videos from Experiment 1 were processed to present the same dyadic interactions with two different display types that varied in contextual information: (1) patch display showing blurred scenes composed of patches (“superpixels”); or (2) body display presenting human body figures on a black background. Participants were asked to rate the degree of threat for each interaction in a display type. Results showed consistency in threat recognition from human interactions regardless of the display type, as the threat ratings in Experiment 2 were strongly correlated with the classification proportion for the raw videos in Experiment 1 (r = .98 for patch display; r =.93 for body display). To examine the underlying psychological dimensions governing threat perception, we used threat rating similarity to conduct the multidimensional scaling (MDS) analysis. The body display MDS result revealed a key dimension that was related to the duration of physical touching. This finding suggests that when contextual information is eliminated, some characteristics of body kinematics serve as critical signals for threat detection.
Acknowledgement: NSF grant BCS-1655300