|
The contribution of top-down predictions to visual recognition| old_uid | 1124 |
|---|
| title | The contribution of top-down predictions to visual recognition |
|---|
| start_date | 2006/04/27 |
|---|
| schedule | 12h45 |
|---|
| online | no |
|---|
| summary | We see the world in scenes, where objects typically appear together in familiar contexts. In spite of the infinitely diverse appearance of these scenes, such context-based associations can give rise to expectations that benefit the recognition of objects within the scene. Building on previous work (Bar, 2003; Bar & Aminoff, 2003), we proposed a mechanism for rapid top-down and context-driven facilitation of object recognition (Bar, 2004). At the heart of this model is the observation that a coarse, low spatial frequency representation of an input image (i.e., a blurred image) is sufficient for rapid object recognition in most situations. Specifically, low spatial frequency image of a scene
can activate a set of associations (i.e., "context frame") that provide predictions about what other objects are likely to appear in the same environment. For example, a 'beach' context frame will include the representation of a beach umbrella, a beach chair, a sand castle, and so on, as well as the typical spatial relations among them. In parallel, a low spatial frequency image of a single target object within the scene can considerably limit the number of alternative interpretations that need to be considered regarding the identity of this object. For example, a low spatial frequency image of a beach umbrella can be interpreted as a beach umbrella, a mushroom, a lamp, or an umbrella. The intersection of these two sources of information, the context and the object alternatives, would result in a unique identification of that object. The resulting representation can then gradually be refined with the arrival of details conveyed by the high spatial frequencies. I will outline the logic and discuss behavioral and neuroimaging data that support various aspects of the proposed model. |
|---|
| responsibles | NC |
|---|
| |
|