Abstract
It is often assumed that multiple object tracking (MOT) experiments can be generalized to real world tracking tasks. However, MOT experiments deliberately use impoverished stimuli; the typical MOT array is a set of identical dots following random trajectories in a plane. In contrast, realistic tracking situations allow observers to exploit knowledge about how objects should move in a given context as well as the distinguishing features of objects, which has been shown to improve tracking performance (Horowitz, et al., 2007; Makovski & Jiang, 2009). In Experiment 1, we asked whether observers can use contextual information to facilitate tracking. We obtained sixty-seven 20 s video clips of basketball games. We then employed tracking software (Ben Shitrit, et al., 2007) to generate matching clips consisting only of moving dots following players’ trajectories. Half of the participants tracked basketball players, the other half dots. There were 2 or 4 targets (equally distributed across teams), designated by flashing green circles for 2 s. After 20 s of tracking, all players or dots were circled in blue and participants marked all targets via mouseclick. Tracking accuracy was superior with basketball videos (p <.05). This could reflect feature information, contextual knowledge, or both. In Experiment 2, we controlled for feature information by using 8-20 s videos of basketball games played normally, reversed, or inverted. The procedure was the same as Experiment 1, except that video condition was varied within-subjects. We found significantly higher accuracy for normal and reversed videos compared to inverted videos (p <.05), suggesting that tracking suffers when context violations make it difficult to predict motion. We conclude that in real world tracking tasks, people exploit contextual information about how objects (basketball players, in this case) move in order to improve tracking performance.
Meeting abstract presented at VSS 2013