High-level perception

What are 3 examples of effortless high-level perception?

1. Identifying objects/ faces

2. Reading words

3. Understanding speech

How fast can we process written words according to Spritz?

400 words/min = 140ms per word

At top speed: 770 words/min = 85ms per word

What is the basic idea of template theories? (Neisser, 1967)

Recognition occurs by matching sensory input to stored templates based on global similarity

What are template theories good at explain?

Simple machine recognition (e.g. bar code readers, ANPR)

What are difficulties with template theories? (3)

1. Completeness issues (e.g. “R” vs “P”)

2. Position, rotation, slant, size variation

3. Difficulties with handwriting and viewpoint changes

What is the core idea of feature matching theories?

Objects are recognised by identifying a set of defining visual features

What is an advantage of feature matching?

A limited set of features can describe a very large number of objects

What did Gibson (1969) propose about features?

Features should efficiently discriminate between alternative with a minimal set

What did Hubel and Wiesel (1962) discover?

Specific neurons in the visual cortex respond to lines at particular angles and positions- simple feature detectors

What is the limitation of viewer-centred representation?

Difficulty recognising objects from different viewpoints

What is an object-centred representation?

Objects are perceived using their own reference frames (axes of elongation/symmetry)

What are “geons”?

Simple 3D shapes used as a basic building blocks for objects

How many geons exist according to Bierderman?

About 24 geons, each with 15 variations

How do geons combine to represent objects?

Through attatchment relations specifying their spatial arrangement

What happens if geon structure information (vertices) is deleted?

Object recognition accuracy drops significantly (70%→ 45%)

What are 2 weaknesses of geon theory?

1. Hard to distingusih objects with similar geon strcutures

2. Poor at explaining recognition of faces and natural objects

What did Biederman (1987) conclude?

That vertices which provide information about geon structure seem critical for object recognition

How do machines achieve object recognition today?

1. Big data training from the internet

2. Deep learning (advanced neural networks)

What is the face inversion effect (Yin, 1969)

Faces are harder to recognise when upside down compared to objects- suggesting holistic face processing

What does the Thatcher Illusion (Thompson, 1980) show?

Configural processing of faces (e.g. nose, eyes, etc.)- inverted faces look odd only when the face is upright