On June 23rd, we discussed “Explaining human multiple object tracking as resource-constrained approximate inference in a dynamic probabilistic model” by Edward Vul, Michael C Frank, George Alvarez, and Joshua B Tenenbaum.
This paper proposes a dynamic probabilistic model as an ideal observer for multiple object tracking. The ideal observer model uses a set of n Kalman filters to track n objects. Observations at each time step are composed of n pairs of position and velocity values. However, inference must be performed in order to associate each observation pair with the appropriate object identity (Kalman filter).
The algorithm for multiple object tracking is accomplished by alternating between 1) a Rao-Blackwelized particle filter to assign identities to each observation and then 2) performing the Kalman filter to analytically track the objects.
This paper goes on to compare the ideal observer performance (the multiple object tracking algorithm) to human confidence during a comparable psychophysical task. The results are comparable, but it would be nice to compare the psychophysical tracking results to the model results directly.
The second focus of this paper was to manipulate the influence of the velocity on the tracking in order to investigate the open question of whether velocity estimates are incorporated into multiple object tracking. By manipulating the linear dynamics that generate the object motion, they are able to evaluate how different assumptions about velocity in the model multiple object tracking compare to human performance. They conclude that the model most closely matches human performance when velocity is not weighted.
This provides a nice algorithmic framework for thinking about how human subjects accomplish multiple object tracking.