Dr. Ben Balas
Determining the features used to make high-level decisions about natural objects, textures, and scenes is a fundamental challenge for vision research. However, the complexity of natural images generally makes it necessary to examine high-level perception with very simplified, unnatural stimuli. This makes certain questions much more tractable, but significantly compromises our ability to generalize results to natural settings. In this talk, I describe an alternative approach that overcomes this difficulty by adopting tools from computer graphics and machine learning to study human perception. Specifically, I will present the results of two studies in which algorithms for texture synthesis were successfully used to examine the nature of texture processing in the human visual system. In the first study, "lesioning" a synthesis model made it possible to determine how the vocabulary of visual recognition differs across distinct categories of natural texture. In the second study, I'll describe an extension of this methodology that has yielded novel insights into the nature of "visual crowding," providing the first general-purpose model of this important perceptual phenomenon.
In both cases, the sophistication of the underlying algorithm makes it possible to ask focused questions about human vision without the need to employ sparse, minimal stimuli. This work serves as an important 'case study' that demonstrates how computer vision methods can be meaningfully applied to human vision research. I close by briefly discussing planned and ongoing extensions of this work within the domain of texture and scene perception, as well as the broader possibilities afforded by this approach for studying high-level recognition.
|
