Purchase this article with an account.
Greg Zelinsky; Combining behavioral and computational tools to study mid-level vision in a complex world. Journal of Vision 2015;15(12):1396. doi: https://doi.org/10.1167/15.12.1396.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
As vision science marches steadily into the real world, a gap has opened between theories built on data from simple stimuli and theories needed to explain more naturalistic behaviors. Can “old” theories be modified to remain relevant, or are new theories needed, tailored to these new questions? It will be argued that existing theories are still valuable, but they must be bolstered by new computational tools if they are to bear the weight of real-world contexts. Three lines of research will be discussed that attempts to bridge this theoretical divide. The first is categorical search—the search for a target that can be any member of an object category. Whereas the largely artificial task of searching for a specific target can be modeled using relatively simple appearance-based features, modeling more realistic categorical search tasks will require methods and features adapted from computer vision. Second, we can no longer simply assume to know the objects occupying our visual world—techniques must be developed to segment these objects from complex backgrounds. It will be argued that one key step in this process is the creation of proto-objects, a mid-level visual representation between features and objects. The role of image segmentation techniques in constructing proto-objects will be discussed. Lastly, the real world creates untold opportunities for prediction. Using Kalman filters, it will be shown how motion prediction might explain performance in multiple-object tracking tasks. Rather than tearing down our theoretical houses, we should first consider remodeling them using new computational tools.
This PDF is available to Subscribers Only