Pattern and object recognition

What is pattern recognition

• A stimulus equivalence problem

• The same object can produce an infinite number of retinal images

• Whereas two different objects can produce the same retinal image

Example of simple recognition mechanisms

• (Tinbergen, 1951)

• Very simple organisms with very simple visual systems

• Aggressive behaviour of male sticklebacks (fish)

• Red belly patch on the stickleback triggers a response

• Even if they have different templates as long as they have the red belly patch this triggers the stimulus

• The detailed model of rival fish is ineffective due to them having entire characteristics as opposed to one distinct characteristic

• Recognition of complex shapes is reduced to one key feature

Examples of sophisticated visual recognition processes

• Eg primates (us), mammals and birds

• We don’t react to simple features

• Requires matching complex stimulus configurations with internal representations

• Pattern recognition of 2D shapes:

• A simplified problem of object recognition

• An applied problem motivated by machine vision

Main approaches of pattern recognition of 2D shapes

• Template theories

• Feature theories

• Structural theories

What are template theories

• Templates of patterns in memory

• Stimulus patterns are compared with templates

• If they match the template, then the stimulus is recognized

Problems with template theories

• We cannot hold an infinite number of templates in our minds.

• Matching would fail even with minor differences between stimulus and template

• Requires standardization processes:

• eg Stimulus- find a major axis- rotate- scale size to appropriate size- scale line weight- template

• Even after standardization, mismatches are likely to occur

• Hard to decide which features to modify to enable matching and what characteristics are essential to identify the stimulus

Experimental evidence against template matching (Sutherland and Williams, 1969)

• Rats are trained to discriminate between A and B

• Transfer test with C and D

• Rats are trained on a A end up selecting C, even if D overlaps better with A

• Rats represent the structure (regular or irregular) of a pattern rather than matching template

• Shows something wrong with template matching approach when looking at biological systems

What are feature theories

• Influenced by Hubel and Wiesel’s discovery of feature detector cells in the primary visual cortex

• Selfridge’s (1959) “Pandemonium” exemplifies many different theoretical models using this approach

• Feature theories would solve the problem of pattern recognition by imagining a letter ‘R’ being projected to the retina and this is just a mix of features

• Mix of features eg vertical lines or curved lines due to us not recognising the stimulus yet

• The cells in the primary visual cortex will respond to one particular type of feature, eg specific cells will respond to vertical lines

• In the next level, some cells will identify a combination of features

Problems with feature detection

• Recognition based on lists of features and doesn’t take into account any relationship between the features

• No information about different instances of a patterns

• This system would recognise two photos as the same thing even if they are different due to them having the same features.

What are structural theories

• More than an approach than a theory

• Roots from Gestalts psychology

• Can be abstract and robust enough to represent patterns of 2D and 3D even if the orientation changes or there are different features

• Eg the letter T, can be a vertical line or a horizontal line (supports or bisects)

• Relationship between the parts is encoded

What is object recognition

• The ability to identify the objects in view based on visual input

• Can look at independent theories (Marr’s or Biederman’s)

• Or dependent theories

Marr’s theory of object recognition

• Complete computational theory

• Mostly bottom up, doesn’t have any assumptions on what the object is

• A modular theory, has different stages which each carry out a different operation but the working of the layers is separate

• Suggests that object recognition is achieved by matching 3D model representations obtained from the visual object with 3D model representations stored in memory as vertical shapes

Key notions in Marr’s approach

• Generalized cones

• Image segmentation and derivation of axes:

• Concavities indicate likely junctions of different parts, then the major axis of each component part can be derived

• One axis has one particular section which is a particular shape however it can change size along the axis

Stages of Marr’s theory of object recognition

• Image segmentation

• Derivation of major axes

• Determining generalized cones

• Matching with 3D templates in memory

Who influenced Biederman’s recognition by components theory

• Influenced by Marr

• Library of 36 primitive 3D shapes “geons” as opposed to generalised cones

• Defined by non accidental properties eg collinearity, symmetry, properties of the shape that don’t change

• Structural descriptions

• Combining geons an infinite number of objects

Recognition by component theory (Bierderman, 1987)

• Regions of most concavity or vertices are more important for identification of objects than mid regions of segments with less curvature (Biederman, 1985)

• The most important regions are those that define relationships between geons (importance of structural descriptions)

• Edge extraction, detection of non accidental properties, parsing of regions of concavity, determination of components and matching of components to object representation

Evidence supporting recognition by component

• (Cooper and Biederman, 1993)

• Participants judge whether 2 objects presented in rapid succession have the same name eg hat

• Longer RT and lower accuracy with geon change than metric change

• Geon change is more disruptive even when metric change is larger than geon change

What is a geon

A basic 3d shape or component used to perceive and recognize objects in the visual field

View invariant approaches to object recognition

• 3D representation of objects

• Predicts equal ease of recognition at most viewpoints

View dependent approaches to object recognition

• Object represented as collection of views

• Predicts easier recognition when objects are seen at those orientations

Evidence for view point invariant representation

• (Biederman and Gerhardstein, 1993)

• The participant is asked to name images of familiar objects as quickly as possible

• Presentation of ‘to be named’ object is positively primed by objects presented in different orientations

Evidence for view point dependent representation

• (Tarr and Bulthoff, 1995)

• Participants were presented with unfamiliar objects such as “greebles”

• Participants familiarized with these greebles in given orientations were slower in identifying them when presented in different orientations

Object categorization and identification

• (Milivojevic, 2012)

• Categorization:

• Concept, eg ‘is this a dog’

• Not slowed down by changes in orientation/viewpoint

• Identification:

• Exemplar, eg ‘is this billy’

• Slowed down by changes in orientation/viewpoint changes from canonical view