Object Recognition Flashcards
The Problem of Universals
How are we able to recognize things in our world (e.g. trees) when all things are different from one another and NOTHING is identical.
Template Models
- Easiest to implement on a computer
- These models are very sensitive & generally fail miserably
Feature-Based Models
- Break down objects (e.g. letters) into component parts & features.
- Identifying common features helps the recognition process
- A finite set of features allow for identification of infinitely variable particulars
Recognition-by-Components Model
- Biederman’s Model: we recognize objects by their 3-D components (called Geons).
- Biederman proposed a maximum of 36 geons
Invariants
Features specific to a given object.
View-Based Models
- We store every view that we see (requires LOTS of memory)
- Recognize by calculating similarity
- Rotate and alter views until you find a match
- Better for face recognitionWorW
Gestalt School
- 1920s
- “Whole is greater than the sum of its parts.”
- Which wholes will be perceived is based on the given set of parts
- The perceptual system uses this to enhance lines
Figure-Ground
While you can switch between the two, you cannot see both simultaneously.
Proximity
Objects that are near each other are more likely to be seen as an organized whole
Similarity
Similar objects are often grouped together
Continuity
We’re more likely to follow continuous lines than try to break them up.
Closure
Objects near a proximity tend to close it off
Symmetry
Symmetrical figures are grouped together rather than separately
Simplicity
The simplest thing is the most likely to occur
Word Superiority Effect
We identify letters better when we have the context of a word (e.g. Rxad = Read).
