High-level perception

Question 1

What are template theories?

Answer 1

• When we recognise something, we match it up with the closest instance (template) of things stored in our memory
• Pattern recognition is based on global similarity match between sensory input and templates stored in memory, the best match is output of recognition process

Question 2

What are the cons of template theories?

Answer 2

• There is a completeness issue, ie for letter recognition, an r could be recognised as a p as it has p in it
• It could be resolves by preprocessing of the visual image, however for handwritten stuff this is an issue

Question 2

Summary of template theories

Answer 2

• It works in some (quite restricted) environments
• For complex objects it needs to deal with changes in viewpoint
• things look quite similar e.g certain animals which makes it difficult

Question 3

What are feature matching theories?

Answer 3

• When pattern recognition is based on identification of features in the visual array
• ‘Features’ are fragments or elementary components of a larger pattern
• Then for recognition purposes, objects can be defined in terms of their component

Question 4

Advantages of feature matching

Answer 4

• A limited nunber of features can be used to represent a very large number of objects
• The features used should be efficient so they should discriminate effectively between possible alternatives, with a minimal feature set

Question 4

What is low-level feature analysis?

Answer 4

• A study conducted single cell recording in the visual cortex of anaesthetised cats
• Specific cells respond only to certain kind of stimuli (e.g a line at particular width angle etc)

Question 5

What is recognition by components?

Answer 5

• Objects can be describes in terms of small sets of geometrical parts named geons - about 24, each with 15 sizes
• geons are simple 3D shapes e.g cylinders, cones, wedges etc

Question 6

How to identify objects through geons?

Answer 6

• Few geons and attachments relations(descriptions of spatial relations among them) can be combines to creates a very large number of objects
• Object recognition relies on attachments between geons otherwise recognition is very difficult
• Vertices seem critical for object recognition

Question 6

What are limitations of geon theory?

Answer 6

• Difficult to distinguish between objects with identical geon structure (e.g horse and cow)
• Recognition of specific individuals (faces) - how can we distinguish between different faces
• Hard for natural objects (etc mountains, puddle), easy for artefacts as they are manmade

Question 7

How do machines tackle object recognition?

Answer 7

• The internet provides many instances on how machines can be trained to be accurate
• deep learning is a novel type of neural networks which show much improved performance compared to previous generations

Question 8

How does face recognition differ from object recognition?

Answer 8

• Faces are more difficult to identify when upside down, whilst upside down objects are much easier to identify
• This suggests that face recognition os more holistic and relies on the whole face, not just individual features