Abstract
Multiple-object tracking has been an influential paradigm for evaluating attention and working memory limits. Recent work has made the consequences of these limits concrete, employing the Kalman filter—a Bayesian state estimator—as a computational model, and demonstrating how noisy estimates of position lead to target/nontarget confusions. Moreover, they have accounted for declining performance with increasing load by assuming that noise in spatial estimates is regulated by limited resources. But these models have also been unrealistic in several ways that we investigated in a series of computational and behavioral experiments. First, extant models sample at the frame rate of presentation. In an experiment with a tracking load of one, we found that models with fast sampling rates (>20Hz) overestimate human performance at high speeds. Second, extant models track all items including nontargets (labeling the targets). This leverages mutual exclusivity to make identity decisions. But it belies the assumption that limited resources impair estimation; if all items are tracked regardless, why does the number of targets matter? When we contrasted target-only and track-all models, the latter did a worse job of accounting for human performance in several ways, including by overestimating performance for tracking with different target loads. Finally, we investigated the possibility that an attentional window might supply an opportunity for mutually exclusive but local identity assignments. We implemented a model that tracked other items only within a limited window surrounding each target. Performance improved considerably compared to no-window models, and fits to human results also improved. Overall, models with limited sampling rates and attentional windows supply a better and more intuitive account of human MOT performance. They also illuminate a wider range of opportunities for a resource-limited observer to respond to load demands, including the control of sampling rate and the scope of attentional selection.
Meeting abstract presented at VSS 2014