L8- Object recognition
Two challenges of object recognition
Recognising the same object under different viewing condition
- This is the problem of object constancy
- It relates also to the inverse problem
Recognising new instances of familiar objects + learning new object categories
Object constancy
We usually recognise the same object (person, face etc) under an infinitely varying
number of viewing conditions
What are viewing conditions?
- Lighting, distance between observer and object, viewpoint, presence of motion,
occlusion, any particular situations affecting the visual system of the observer etc
Size constancy
, Each glance at an object generates a retinal image that is almost always different from
any other glance you’d have taken at the same object
What makes a good viewpoint? (1)
Maximises the amount of non-accidental information
Reveals the 3D structure of the object
Includes diagnostic object parts + features
Non-accidental viewpoint- rotating the object around the view does NOT dramatically
change its appearance
- When it does change- the viewpoint is ‘accidental’
What makes a good viewpoint? (2)
How should we interpret the variation in goodness-of-view ratings?
- Ratings are not very informative
Are intermediate viewpoints equally good/bad?
- Some viewpoints are BUT some aren’t
- It also depends on the object
Paired-comparison experiment
Two challenges of object recognition
Recognising the same object under different viewing condition
- This is the problem of object constancy
- It relates also to the inverse problem
Recognising new instances of familiar objects + learning new object categories
Object constancy
We usually recognise the same object (person, face etc) under an infinitely varying
number of viewing conditions
What are viewing conditions?
- Lighting, distance between observer and object, viewpoint, presence of motion,
occlusion, any particular situations affecting the visual system of the observer etc
Size constancy
, Each glance at an object generates a retinal image that is almost always different from
any other glance you’d have taken at the same object
What makes a good viewpoint? (1)
Maximises the amount of non-accidental information
Reveals the 3D structure of the object
Includes diagnostic object parts + features
Non-accidental viewpoint- rotating the object around the view does NOT dramatically
change its appearance
- When it does change- the viewpoint is ‘accidental’
What makes a good viewpoint? (2)
How should we interpret the variation in goodness-of-view ratings?
- Ratings are not very informative
Are intermediate viewpoints equally good/bad?
- Some viewpoints are BUT some aren’t
- It also depends on the object
Paired-comparison experiment
