Why is object recognition difficult (2 reasons)
- objects overlap but form coherent perceptual experience
- retinal image varies but size and shape of object doesn’t
Challenges of object recognition: variability on the retinal image
translation, rotation, size, colour
Challenges of object recognition: Variability of a visual scene
occlusion and presence of other objects
challenges of object recognition: intra-class variation
an object is still identified as such despite it being a different shape, colour, design (features). Eg diff types of chair. Can also be recognised when partially hidden.
Challenges of object recognition: viewpoint variation
can identify objects regardless of what view were seeing them from
2D pattern matching theory: template theories
- template in LTM of all known patterns. Can have multiple templates and compare for greatest overlap
- eg barcodes and fingerprints
Issues with template theory
imperfect matches and not accounting for flexibility of pattern recognition system (need to be identical orientation, size, position etc)
2D pattern matching: prototype theories
- more flexible templates
- process the average of characteristics
- no perfect match
Evidence for prototype theory pattern matching
Franks and Bransford 1971: presented object based on a prototype, Ps were confident they’d seen the prototype before seeing the images that made the average prototype
2D pattern matching: feature theories: what is it and a limitation
- pattern has set of features/attributes
- must know the relationship between the features
- too many objects for this to be the only mechanisms
2D pattern matching: structural descriptions
- a breakdown of the pattern and relationship between it’s features
3D object recognition
comprised of early image processing: objects separated from background. segregated from each other and then matched to object description in memory
Marr and Nishihara 1978: 3D object recognition theory
- the relationship between objects made up of cylinders is it’s structural description
- not possible for all objects, good for biological things
3D object recognition: Biederman’s recognition by components theory
- objects composed of geons (geometrical ions)
- 36 diff volumetric shapes
- concave parts of object can help segment image into separate geons
This theory is viewpoint invariant
Recognition by Components theory (Biedermans geons)
Marr and nishiharas cylinders
Properties to decide what geons make up an image (5)
Curvature, parallel, co-termination, symmetry, collinearity
Evidence for recognition by components theory: concavity (Biederman 1987)
taking away the points of concavity made it harder to recognise as the geons couldn’t be identified
Evidence for recognition by components theory: cortical neurons in monkeys (vogels et al 2001)
some neurons in inferior temporal cortex responded more to changes to geons than object size changes
Disadvantages/evaluation of recognition by components theory
- Edge information/structure not always key: texture and colour also important
- ignores viewpoint dependence: easier to recognise some objects from some views
- ignores within category discrimination (all chairs dont look the same)
Viewpoint invariant theories are better for easier between-category discrimination. Viewpoint dependent theories are better for…
harder between-category discrimination and within category discrimination
Define viewpoint dependence:
Some objects are faster and more accurately defined from certain viewpoints
The binding problem asks:
how do we integrate different kinds of information when identifying objects?
What are the stages of object recognition according to Humphreys et al 1988
- separation of object from background/other objects
- structural description
- semantic representation: ie how is it used
- name representation
Agnosics struggle with which part of object recognition?
structural description
