Lecture 3 - Object Recognition Flashcards
What’s the diff for humans + computers perceiving objects?
H: perception of familiar items
C: perception of familiar patterns
Why is object recognition difficult?
Environment has hundreds of overlapping objects but perceptual experience structure + coherent (we use/name)
Apparent size/shape of object doesn’t change despite large variation in retinal image
What are 4 types of variability in recognition?
- Translation invariance (diff locations in field of view)
- Rotation invariance
- Size invariance
- Colour invariance
Other factors in object recognition?
Partial occlusion, presence of other objects
Intra-class variation (variety of chairs that don’t look the same)
Viewpoint variation
4 types of theories of 2D pattern matching?
- Template
- Prototype
- Feature
- Structural
What are template theories?
Mini-copy/template in LTM of all known patterns, multiple templates in memory then compare stimuli to templates to find greatest overlap
Problem with template theories?
Normalisation? Numerous templates? (examples: barcodes, fingerprints)
Problem: imperfect matches, flexibility of pattern recognition system, comparison requires identical orientation/size/position to stimuli
What are prototype theories?
Modification of template matching (flexible templates), possesses average of each individual characteristic, no match perfect, criterion for matching is needed - Franks/Bransford study
What are feature theories?
Pattern consists of set of features/attributes (A = 2 straight lines + connecting bar) –> but also need to know relationship between features? / \ = A??
What are structural descriptions?
Describe nature of components of configuration + structural arrangement of these parts
How is 3D object recognition more difficult?
First interpret input to visual system as coherent structures, segregated from one another + background –> processed to give description which can be matched to description in memory
What did Marr + Nishihara study?
Objects made of cylinders, must specify relationship between cylinders to make structural description (hierarchical organisation of cylinders)
What did Biederman do?
Alternative to Marr + Nishihara – Recognition by components theory: objects composed of basic shapes (36 GEONS – geometrical ions) viewpoint invariant theory, use structural relations between the parts
How are geons specified?
Non-accidental properties:
Curvature (points on curve), parallel (set of points in parallel), co-termination (edges terminating in common point), symmetry (vs asymmetry), co-linearity (points in straight line)
What did Biederman test (forms of degredation)
Forms of degradation which disrupt basis for identifying geons should make objects more diff to recognise – deleted edges at points that were easily reinstated/diff to determine
How did deletion of components affect recognition?
Slow/inaccurate at non-recognisable but relatively good at recognisable
Deletion of component affects matching stage, midsegment deletion makes it more diff to determine components –> at brief exposure partial objects better recognised but longer midsegment deletion led to less errors
Evaluation of Biederman’s Geons?
Good: recognises importance of arrangement of parts, parsimonious (small set of primitive shapes)
Bad: structure not always key to recognition, which geons?, within category discrimination, de-emphasise role played by context, simplifies contribution of viewpoint-dependence, experiment consistent but not critical test
What is viewpoint dependent theory?
Assume changes in viewpoint reduce speed/accuracy of object recognition
Object representations collections of views that depict appearance of objects form specific viewpoints
Under what circumstances is viewpoint invariant/dependent theories important?
Dependent: complex within category
Invariant: easy categorical
What is Humphreys model of recognition?
Object –> structural description –> semantic representation –> name representation –> name
Problem with object recognition theories?
Oversimplification, later processes might start before earlier ones completed
Support from patients w/ object recog diffs (associative agnosia) - patient HJA, JB
JB - naming visually confusing objects had knock on effects, diff to identify category
Alternative to Humphreys mode?
Cascade model - structual/semantic/name stages interact within/between stages
What are agnosias?
Failure of knowledge/recognition
Visual agnosia: feature processing/memroy remain intact, recognition defivits limited to visual modality
Touch/smell may substitute in recognition
What is apperceptive agnosia?
Problems with early processing (shape extraction)
Perceptual deficit affects visual representations directly, components of visual percept picked up but can’t be integrated –> unusual views of objects
What is associative agnosia?
Problems w/ later processing (recognition), visual representation intact but can’t be accessed/used in recognition, lack of info about percept
