object recognition Flashcards
1
Q
perception
A
- Perception refers to our ability to extract meaning from sensory input.
- It includes audition, taste, touch and olfaction (smell), but research is dominated by vision.
- Vision alone accounts for over 50% of all neurons in our cortex.
- How many senses do we have? - 9-22
- PERCEPTION IS A CONSTRUCTIIVE PROCESS!!!!!
2
Q
visual system
A
- Brain receives sensory input allowing us to perceive an object.
- Cognitive system “constructs” perception.
3
Q
object recognition - a 3 stage model
A
- Image
- Stage 1 - local features - edge detection/contrast, etc
- Stage 2 - shape representation - Gestalt principles/feature integration
- Stage 3 - object representation - stored representations/knowledge.
4
Q
gestalt principles
A
- The whole visual percept is more than the sum of its parts.
- Our perceptual system constantly tries to impose organisation on its input.
- Components of an image are grouped together on the basis of certain visual properties.
- Laws of perceptual organisation e.g., “good continuation” and “closure” give rise to “illusory contours”.
5
Q
shape perception
A
- Primarily “bottom up” processes produce a “primal sketch” (Marr, 1982).
- This sketch contains “primitives” - edges/ orientations/ positions/ lengths/ colours etc.
- “Top down” processes (e.g., Gestalt laws) are used to group collections of primitives together into “lines, curves, larger blobs, groups and small patches” - symbolic primitives.
6
Q
object recognition - 3 models
A
- Template matching (prototype theory).
- Feature analysis
- Recognition by components (structural theory)
7
Q
template matching
A
- A template is an internal representation.
- A memory against which the visual input is matched.
- Computer based object recognition programmes use templates.
- Intuitively plausible - object recognition must involve some kind of contact with a “comparable internal form”.
- But what rules determine whether a match is made?
- How many different templates are needed?
8
Q
feature analysis
A
- Assumes lower level features are analysed first.
- The perceptual system searches for simple characteristic features of an object.
- Supported by neurological evidence (e.g., orientation selective cells in visual cortex).
- Most research focusses on letter/ word recognition.
- Letter A made up a / \ and -
- But what about V and X? both made of / and ... Spatial relationship importants.
- What about complex natural objects?
9
Q
recognition by components
A
- Any specific view of an object can be represented as an arrangement of simple 3D shapes - geons (Biederman, 1987)
- Geons are “viewpoint invariant” - easily “recoverable” from a 2D retinal image.
- “Invariant properties” include cotermination and parallelism.
- Object recognition is impaired when geons are made “non-recoverable” by removing termination points.
10
Q
summary
A
- Object recognition is rapid and apparently effortless.
- It is a constructive process, not a passive one.
- It combines “bottom up” processes (e.g., edge detection) with top down processes (e.g., expectations/ memory).
- A number pf models of object recognition have been put forward.
- These models are not necessarily mutually, exclusive, and object recognition may well involve elements of template matching/ feature analysis/ recognition by components.
- Different strategies may be used in different situations.