Abstract
Intuitively, extrapolating an object’s trajectory should facilitate visual tracking. This intuition has motivated several investigations into whether humans extrapolate when tracking multiple objects. Surprisingly, they appear not to, largely ignoring an object’s motion, and implementing a low-level, spatial memory function instead. Why do humans appear not to extrapolate if extrapolation is beneficial? We address this question by interrogating its underlying assumption: that extrapolation is beneficial in the first place. Recent advances have characterized tracking mechanisms as implementing correspondence operations over noisy inputs of object positions, speeds, and bearings. With noisy inputs noisy predictions should emerge. What is the marginal advantage of extrapolation given noisy predictions? We used a Kalman filter model to answer this question, endowing it with perceptual limits in the range of typical human observers. Specifically, we tested the model with a range of spatial precision limits as well as a range of limits on how quickly it could sample inputs. We compared the model’s performance to a nearly identical model that does not extrapolate. In Experiment 1, we found worse performance for a model that extrapolated relatively rigidly compared to one that did not extrapolate at all. In Experiment 2 we found negligible advantages for a model that learned to weight its own predictions, and that the model placed little weight (<15%) on extrapolation. Finally, Experiment 3 demonstrated that a simple model that does not extrapolate replicates a signature finding in the literature on human tracking: improved performance following a global interruption for objects that remain stationary, compared with objects that persist along stable trajectories. Together, the results explain the finding that humans do not seem to extrapolate when tracking multiple objects. In particular, given human perceptual limits, extrapolating from noisy inputs does not confer a marginal advantage.
Meeting abstract presented at VSS 2013