| A key challenge for vision research is to understand how recognition works in complex, natural environments. To that end, I've studied recognition behavior in multiple settings. Each one of the areas discussed below offers an important opportunity to study a crucial aspect of visual behavior in our cluttered and ever-changing visual world. Relevant publications are highlighted alongside each topic's summary. Face Recognition Faces are perhaps one of the most important visual stimuli we encounter in our environment. They're also an excellent example of an "overlearned" or "expert" object class. Typical observers can extract a wealth of information from a face in a fraction of a second, including who it is, how they're feeling, and whether or not they're attractive. How does the visual system do it? Face recognition is a valuable "case study" of expert-level recognition and how useful visual features are learned, applied, and refined over time.  A computer-generated face (A) and a "Cartoon" (B) rendered from it for use in ongoing behavioral experiments. In previous work, I've asked observers to create "portraits" of individuals by placing pre-segmented face parts into an outline (C) as a means of investigating the fidelity of so-called "configural" or "2nd-order" representations of facial appearance. In my work, I've explored computational strategies for representing faces for recognition and conducted behavioral and electrophysiological studies of observer's abilities. An overarching goal for me is to define previously-suggested face processing mechanisms in terms of explicit computational algorithms. For example, what does it mean to say that a face is recognized "configurally" or "holistically?" What consequences do these mechanisms have on other aspects of recognition, and how do they emerge during learning and development? Texture Perception and "Statistical Vision" Typical observers can recognize both the "things" and the "stuff" around them, a dichotomy which refers to the visual world's constituent objects and materials respectively. While there is a wealth of data describing observers' capabilities and limitations in the domain of object recognition, there is comparatively little work exploring how surfaces, textures, and materials are recognized. This represents a distinct frontier for research in high-level vision, especially since the representation of a texture is likely to be "statistical," incorporating measurements of feature distributions rather than maintaining highly localized pattern information.  Using texture synthesis models makes it possible to determine what statistical information is actually used to make various visual decisions regarding texture properties. This picture shows how texture "lesioning" can be used to isolate particular features and make images that differ from a parent texture (coffee beans, at top) in very specific ways. How discriminable these images are from the parent texture says a lot about the underlying visual representation. My research explores the nature of "statistical vision" through the use of models for texture synthesis originally developed for use in the computer graphics community (see above). These models offer a powerful vocabulary for expressing the features used for "Statistical" recognition in textures, scenes, and most recently, "crowded" displays. What are the limits of such a representation of visual structure? How much can we do without measuring particular patterns at particular locations? Recognizing Moving Objects The world around constantly moves and changes. We walk through our environment, constantly moving our eyes from one changing object to the next. From a computational perspective, this presents an almost intractable level of complexity. How are objects segmented, tracked, and recognized given the speed with which our visual world changes? We not only cope with an ephemeral visual world, we use the information gathered over time to make additional judgments about what is going on around us. We can identify people's stride as masculine or feminine and recognize our friends from a characteristic gesture or expression, all in a startingly short time.  Does the way an object moves play a role in determining how we remember what it looks like? In my thesis work, I examined how motion contributed to memory for novel "paperclip" objects that spun around their vertical axis during object learning (A). Observers then attempted to recall particular "frames" they saw during this initial exposure and I attempted to model their performance using a model that lacked any information regarding motion (1st panel at right) and also using a model that did incorporate such information (2nd panel at right). Knowing how an object moves dramatically improved model performance, suggesting people use object motion, too. In my thesis work, I explored how observers use the observation of object motion to learn about static form. I found that observers use motion to update and refine their representation of novel shapes, providing them with a useful representation supporting both tolerance and sensitivity to appearance change. My current efforts are directed towards understanding the nature of dynamic object representations. Many basic questions have yet to be answered regarding moving objects: How do representational strategies change as speed increases? How invariant are moving objects to extrinsic and intrinsic variability? What is the time scale for processing dynamic visual features? |