Experimental goal Experimental Paradigm Results Illusions of nonrigidity

Results (contd. ) 10% X-Y positional noise added 3 x 10% 2 x 20%

Inferences 1. Arbitrary associations between 2 D and 3 D structures can be learned.

A model for incorporating high-level learning in early perception (Jones, Sinha, Poggio, Vetter, Current

The model schematically Learned knowledge about object-class n Learned knowledge about object-class 1 Image

Combining learned instances Object class Image 0 + Attribute 0 Image 1 + Attribute

The model in action: Example 1 - implicit 3 D shape recovery Image 45

The model in action: Example 2 - Clean edge-map extraction Image frontal face image

Example 2 (contd. ) Strengths of high-level learning based model: 1. Ability to complete

Slides: 16

Download presentation

Experimental goal

Experimental Paradigm

Results

Illusions of non-rigidity

Results (contd. ) 10% X-Y positional noise added 3 x 10% 2 x 20% 1 x 20% Training set 30% 100% 30% 75% 40% 50% 40% 25% 50% 0% 50% Training set Percentage of trials over which subjects reported non-rigidity Learning is object-specific. . . Scale factor

Inferences 1. Arbitrary associations between 2 D and 3 D structures can be learned. 2. Learning is object-specific. 3. Learning influences perception of kinetic depth and stereoscopic depth. 4. Recognition may, in some circumstances, precede 3 D shape perception. Recognition 3 D Shape Image

A model for incorporating high-level learning in early perception (Jones, Sinha, Poggio, Vetter, Current Biology, 1997)

The model schematically Learned knowledge about object-class n Learned knowledge about object-class 1 Image 1+attribute Image 2+attribute … Imagen+attribute Attribute estimation via prototype combination Result Recognition (somehow…) Image

Combining learned instances Object class Image 0 + Attribute 0 Image 1 + Attribute 1 Imagen + Attributen image_new + ? Step 1: image_new = Step 2: + C 0 * Image 0 C 1 * Image 1 + Cn * Imagen C 0 * Attribute 0 + C 1 * Attribute 1 + Cn * Attributen = Attribute_new

The model in action: Example 1 - implicit 3 D shape recovery Image 45 degree face image Attribute Frontal face image