Multiple object tracking difficulty accounted for by an ideal observer

摘要

Our ability to track multiple targets among distractors is influenced by the number of objects onscreen, as well as their speed, spacing, dynamics, and superficial features. Here, we compare human performance to an ideal object-tracking observer to predict individual trial difficulty and the influences of tracking manipulations. We measured the maximum object speed at which observers were able to maintain 75% accuracy while varying object spacing, number of distractors, object color fidelity, tracking duration, and motion smoothness. Additionally, we measured the performance of 250 observers on 100 fixed trial trajectories (across-trial variation in accuracy was highly reliable over observers, r=0.63). We then compared trial accuracy and speed-spacing tradeoffs to the ideal observer tracking under the same manipulations. The ideal observer tracks each object via a Kalman filter, conditioned on the correspondence of represented objects to onscreen objects (sampled via a particle filter). We fit one parameter corresponding to noise in perceived spatial position to achieve 75% accuracy averaged over the measured speed-spacing thresholds. With no further fitting, the same model captures the variation in difficulty across trials (correlation of model to observer accuracy: r=0.50) as well as across task manipulations. Notably, the model explained the counterintuitive effect of better performance in conditions with less predictable motion. Although we could capture much of the reliable variation in trial difficulty, the ideal observer failed to capture the pattern of errors within a trial. This suggests that the same resource limitations that are necessary to account for effects of the number of targets may also underlie the systematic object errors. In summary, we show that an ideal object-tracking observer provides a good account of the changes in performance due to task parameters like speed, spacing, number of distracters, and object motion, and also the reliable variation in difficulty across trials.