Maria Kon¹ (), Sangeet Khemlani¹, Andrew Lovett¹; ¹U.S. Naval Research Laboratory

Chase detection is a demanding task in which participants view a scene with moving objects and indicate whether one is chasing another. While this task offers insight into intention attribution, the mechanisms of chase detection remain poorly understood. Recent work (Kon et al., 2024) introduced a new chase detection paradigm that measured response time and accuracy while varying set size and whether a chase was present or absent. The authors developed an attention-based computational model that made specific predictions about performance. Here we present the results of an experiment that used the same chase detection paradigm but also varied a cue introduced by Gao et al. (2009), i.e., chase subtlety (the degree to which the chasing object deviates from a direct, “heat-seeking” path towards the chased object). Prior to collecting data, we conducted model simulations and made preregistered predictions about performance. Those simulations correlated well with the data, with two notable discrepancies: (1) human response times for chase-absent trials were much lower compared to the model; (2) for chase-present trials with the highest chasing subtlety, the model had lower accuracy and faster response times compared to humans. In light of (1), we refined the model’s stopping rules. (2)’s mismatches may be the result of human participants learning to classify videos with higher chasing subtlety as chases due to feedback received after each trial, and they may have waited longer to determine whether there was such a chase. We tested these hypotheses by conducting a follow-up experiment that provided no feedback on experimental trials. Thus, the model both makes testable predictions about task performance and suggests the need to reassess the role of chasing subtlety in such tasks. Further, the experimental results provide an opportunity to refine the model, leading to a deeper understanding of how humans detect chases.

Acknowledgements: This research was supported by a National Research Council Research Associateship awarded to MK and by a grant from the Naval Research Laboratory awarded to SK.