CSE259 Nicholas Butko
Infomax Control: Closing the Perception-Action Loop
While it is tempting to think of Eye Movements in terms of 'Information Gathering', it is not immediately apparent how to measure the information gathered by a single fixation. How we choose to measure the information content of a visual scene has big consequences on the kinds of eye-movement models that result. In this talk, we contrast two different eye movement models both grounded in different notions of information gathering. The first model, visual saliency, arises from viewing the image itself as a series of events generated by an information channel. The second model, the I-POMDP model of visual search, arises when the image is seen as containing information about the answer to a question. The resulting algorithm can call any object detector as a subroutine, and increases the speed of object detection. For example, we increase the speed of the OpenCV Viola Jones face detector by a factor of two.