Abstract
Detecting visual changes in environments is an important computation that has applications both in the study of human and computer vision. Additionally, task oriented descriptions of visual cognition [1] could use such a mechanism for switching between ongoing tasks and updating internal visuospatial memory representations. We conjecture that the number of environments in which people spend most of their time is limited (out of the set of possible visual stimuli), that the environments do not frequently undergo major changes between observations and that over time subjects learn distributions of spatial and visual features. These assumptions can be exploited to reduce computational complexity of processing visual information by utilizing memory to store previous computations. A change detection technique is then required to detect when predictions deviate from reality. Baldi and Itti [2] provided a Bayesian technique, however they did not incorporate spatial location of features and require a commitment to a distribution of features a priori. We propose a mechanism that instead uses a low dimensional representation of visual features to generate predictions for ongoing visual stimuli. Deviations from these predictions can be rapidly detected and could be used as a reorienting attentional signal. The model we present is computationally fast and uses a compact descriptions of complex visual stimuli. Specifically, the model encodes color histograms of naturalistic visual scenes captured while exploring an environment. It learns a spatial layout of visual features using a self-organizing map on location data and color data, compressed using the matching pursuit algorithm. We present tests of the model on detecting changes in a virtual environment, and preliminary data for human subjects' change detection in the same environment.
[1] Sprague and Ballard, Proceedings of the 18th IJCAI, August 2003
[2] Baldi and Itti, ICNN&B 2005.