Abstract
Research on multiple-object tracking suggests that the visual system can track targets through occlusions by extrapolating future positions from past motion. Evidence for such extrapolation is clearer with a smaller number of targets. Specifically, Luu and Howe (2015) showed that participants were able to track two targets better with predictable motion trajectories than with unpredictable ones. In contrast, accuracy was unchanged across conditions when tracking four targets unless trial types were presented in blocks. To investigate extrapolation during tracking, we developed a computational model based on two ideas. First, a pre-attentive, parallel process tracks targets by enhancing their locations while inhibiting the surrounding area. Second, a post-attentive, serial process computes motion trajectories for a single target and predicts its future location. Attention is drawn to targets that are crowded by other objects. As the number of targets increases, so does the amount of crowding and the number of occlusion events. As a result, the ability to serially extrapolate each object's location through occlusions decreases. Following their procedure, we compared the model to Luu and Howe's experiment 1. In line with the human data, the model's accuracy was higher in the two-target condition when trajectories were predictable (81.1% vs. 68.3%, p< .05). Contrary to their findings, accuracy was also higher in the four-target condition with predictable trajectories (81.6% vs. 72.8%, p< .05). These results suggest that although the model struggles with four targets—to equate accuracy, videos with four targets must be run at half the speed—it still benefits from motion extrapolation. Luu and Howe's experiment 4, where unpredictable trials appeared in separate blocks, similarly showed extrapolation for both two and four targets. Perhaps people rely on extrapolation in easy (two targets) or reliably beneficial (blocked conditions) situations, whereas the model always performs the more effortful strategy.
Meeting abstract presented at VSS 2017