Lecture 3, Object Recognition Flashcards
What is object recognition?
Perception of objects is different for humans and computers.
For humans –> perception of familiar items
For computers –> perception of familiar patterns
Why is object recognition difficult?
- Environment contains hundreds of overlapping objects
- Objects have variability, e.g., translation invariance, rotation invariance, size invariance, colour
- Objects can vary in the visual scene, e.g., partial occlusions and presence of other objects
- Intra-class variation -> same object has different forms, e.g., different types of chairs
- Only part of object may be visible
- Viewpoint variation -> may be more difficult to recognise an object from one viewpoint to another
What are theories of 2D pattern matching?
Template theories, prototype theories, feature theories, structural descriptors
What are template theories?
Mini copy/template in LTM of all known patterns - compare stimuli to templates in memory for one with greatest overlap until a match is found
What are some problems with template theories?
- Imperfect matches
- Cannot account for the flexibility of pattern recognition system
- Comparison requires identical orientation, size, position of template to stimuli
What are prototype theories?
Modification of template matching (flexible templates) - possesses the average of each individual characteristic. No match is perfect –> criterion is missing
Is there any supporting evidence for prototype theories?
Documented by Franks and Bransford:
–> Presented objects based on prototypes
–> Prototype not shown
–> Ppts confident they had seen prototype
–> Suggests evidence of prototype
HOWEVER, difficult to conclude how you would come up with a prototype
What are feature theories?
Pattern consists a set of features or attributes, e.g., letter A = 2 straight lines and connecting bar. However, also need to know the relationship between features. e.g., / \ - = A?
What are structural descriptions?
Describe the nature of the components and the structural arrangement of these parts. Composition of object and how they are related together. e.g., Capital letter T = 2 parts; 1 horizontal, 1 vertical; vertical support horizontal; vertical bisects horizontal.
What is 3D object recognition?
Must interpret input to the visual system as coherent structures, segregated from one another and from background. Must be processed to give a description which can then be matched to the descriptors of visual objects stored in memory.
What is Marr & Nishihara’s theory of 3D object recognition?
Objects are comprised of cylinders; must specify relationship between cylinders (structural description). Expressed structural relationships by a hierarchical organisation of cylinders. Each cylinder has an axis and way in which others are joined are expressed as coordinates.
What are the limitations of Marr and Nishihara’s theory of 3D object recognition?
–> Difficult to think about how to break down all objects into a series of cylinders
–> May work better for biological entities
What is Bierderman’s theory of 3D object recognition?
Provided an alternative to Marr & Nishihara’s theory.
- Objects comprised of basic shapes
- GEONS -> Geometrical ions
- Approx 36 different shapes
- Viewpoint invariant theory
- Relationship between geons can be described structurally
What are some examples of the structural relationships set out by Biederman?
Relative size, verticality, centring, relative size of surfaces at join.
What did Biederman believe concave parts of objects were helpful for?
Segmenting visual image into parts
Biederman specified GEONS in terms of “non-accidental” properties, what were they?
- Curvature: points on a curve
- Parallel: sets of points in parallel
- Co-termination: edges terminating in a common point
- Symmetry: vs. asymmetry
- Co-linearity: points in a straight line
What does “non-accidental” mean?
Regularities in the visual image thought to reflect regularities in the world.
2D symmetry in the visual image indicates symmetry in 3D object -> present in both 2D & 3D views
What did Biederman say about form degradation?
It should disrupt the basis for identifying geons; make object more difficult to recognise; deletion of component affects matching stage, reducing the number of components to match to.
Midsegment deletion makes it more difficult to what?
Determine components
Describe Biederman’s study that documented effects of degradation?
-> Deleted edges at points where easily reinstated or difficult to determine
-> 25, 45, or 65% of contours removed
-> Slow & inaccurate at “non-recognisable” but relatively good at “recognisable”
-> At brief exposures -> partial objects are better recognised
-> At longer exposures -> midsegment deletion led to less errors
What did Vogels, Biederman, Bar & Lorincz, 2001, find in monkeys sensitive to geons?
- Cortical neurons.
-> Assessed response of individual neurons in the inferior temporal cortex to change in geon or change in size of object.
-> Some neurons responded more to geon changes, providing support for geons
What are the evaluations of Biederman’s model?
- Flexible and comprehensive system for describing objects, but why 36 geons?
- Experimental results consistent with model but doesn’t provide critical test
- Doesn’t explain how description are matched to those stored
What are the advantages of Biederman’s theory?
- Recognises importance of the arrangement of the parts
- Parsimonious: small set of primitive shape
What are the disadvantages of Biederman’s theory?
- Structure is not always key to recognition, e.g., Peach vs. Nectarine -> texture is necessary
-> Within category discrimination
-> De-emphasise the role played by context in object recognition (affects later stages of object recognition)
-> Simplifies the contribution of viewpoint-dependence
What do viewpoint-dependent theories assume?
Changes in viewpoint reduce the speed and/or accuracy of object recognition. Object representations are collections of views that depicit the appearance of objects from specific viewpoints.
What is viewpoint dependent more important for?
Within-category discrimination
What does evidence suggest concerning viewpoint invariant mechanisms vs viewpoint dependent mechanisms?
Viewpoint dependent = complex within category decisions
Viewpoint invariant = easy categorical decisions
What is Humphrey’s Cascade Model?
Structural, semantic and name stages - both within and between stages.
-> Makes different predictions about how subjects will perform in object naming task.
–> Problems at one stage will have a knock on effect.
e.g., Patient JB - naming visually confusable words (birds, animals) had knock-on effects, making it difficult to identify their category
What is agnosia?
Failure of knowledge or recognition
What is visual agnosia?
-> Feature processing and memory remain intact
–> Recognition deficits are limited to visual modality
—> Alertness, attention, intelligence and language are all unaffected
—-> Other sensory modalities (touch, smell) may substitute for vision in allowing objects to be recognised
What is apperceptive agnosia?
Problems with early processing (shape extraction)
-> Perceptual deficit, affects visual representations directly
–> Components of visual percept are picked up, but can’t be integrated
—> Effects may be graded
—-> Often affected: unusual views of objects
What is associative agnosia?
Problems with later processing (recognition). Visual representations are intact, but cannot be accessed or used in recognition - lack of information about the percept. “Normal percepts stripped of their meaning.”