lecture 2- object perception Flashcards
1
Q
how are the same object at different perspectives processed?
A
- recognise objects as same
- but visual info coming into retina is differet
2
Q
what is meant by gestalt psychology?
A
- ‘grouping principles’ of perceptual organisation
- group things in mind depending on different criteria
- parts of image seen as belonging together (these parts are likely to arise from the same object)
- ‘whole is more than the sum of parts’
3
Q
what are the gestalt grouping principles
A
- similarity (e.g in luminance, shape, colour)
- proximity (tend to see groups where closer together)
- closure (closing areas of space into e.g a square)
- good continuation
- common fate (discs moving together at same speed and direction vs. still ones)
4
Q
what is a figure-ground?
A
- area bounded by contour (closure) is seen as a separate object
- contours seen as belonging to one object at a time (can only see one option at a time, not simultaneously)
5
Q
what is Marr’s model of recognition?
A
- primal sketch of what objects look like- 2D rep of luminance (enables us to detect edges and contours)
- 2 1/2 D sketch- description of depth, orientation, shading, texture, motion, binocular disparity- (viewpoint dependent)
- 3D model- description of 3D shape of object (view point invariant)
- analyse image with range of edge filters (hubel weisel cells in visual cortex)
- use gestalt grouping (continuity) to find outline
- segment outline at concavities (dips), from there identify principal axis of sections
- define arrangments of parts (cylinders): start with biggest (principal axis) and work progressively through smaller
NOTE: visibility (angle) of principal axis is important
6
Q
explain Biederman’s ‘recognition by components’ computational model
A
- edge extraction- surface characteristics: luminance, texture, colour
- detect arrangment of edges: curvature, parallel, co-terminating, symmetry (don’t alter with viewpoint)
- segment object into components: detect concave parts
- determine GEON type for each component: 36 needed- any combo of these create any object (most only need few)
- determine arrangment of GEONs and match GEON description to memory
7
Q
what are the problems with Biederman’s model?
A
- does not differentiate objects within class (telling difference between different faces/mugs)
- does not use surface pattern
- recognition is viewpoint invariant (can recognise regardless of viewing angle) BUT can matter when see object from
8
Q
what are the two object processing pathways? (Ungerleider & Mishkin 1982)
A
- ‘what’- bilateral removing of inferior temporal area (TE) = severe impairment in object discrimination, but can grab correct option well
- ‘where’- bilateral removal of posterior parietal cortex = impairment in landmark discrimination, but could do novel object discrimination
- DOUBLE DISSOCIATION
9
Q
what are the symptoms of object agnosia and which pathway does it affect?
A
- ventral visual pathway
- no loss of intelligence
failure to recognise objects - no simple visual impairment
- can see edges but cannot put them together
- may draw object ok but not recognise drawing
10
Q
what is the case study for object agnosia?
A
DF- CO poisoning, Milner’s posting task
- can: accurate guidance of hand and fingers toward object (can reach/grasp/pick up)
- can’t: object size/shape/orientation (when asked to indicate size with fingers/match angle of object with slot when holding infront of it)
11
Q
what is the ‘titchener circles’ illusion? (Aglioti 1995)
A
- target circle size is influenced by surrounding array of circles
2 options: - discs physically the same (but look different)
- discs perceptually the same (but actually different)
- perception of disc size strongly affected by illusion BUT grip aperture correct despite it
- shows object analysis for action separate from conscious perception of object
12
Q
what are the symptoms of optic ataxia and which pathway does it affect?
A
- dorsal visual pathway
- difficulty completing visually-guided reaching tasks
- difficulty reaching in the right direction
difficulty positioning fingers correctly towards an object - little relationship between grip aperture and object size
13
Q
what are views on the purpose of vision?
A
- early models: to construct an internal model of reality (outside = reality, trying to create accurate version of outside in head)- foundation for all visually derived thought and action
- now: more focus on ‘requirements of vision’ (Goodale & Milner 1992)- considering what visual info is doing for us (instead of making neural representation- modularity based upon what ‘uses’ vision can have)
14
Q
what are the requirements of vision?
A
- vision for perception (identification of object- ‘object-centered’)
- vision for action (how we can change that thing in space- ‘viewer-centered’)
- require different transformations of visual signals
- need to encode size, orientation and location of objects relative to others
- start with this frame of reference- needs a perceptual representation of objects that transcends particular viewpoints (while preserving info about spatial relationships)
15
Q
what is the function of the LOC?
A
- lateral occipital cortex
- how we recognise the same object in more than one location => identity representation
- location-tolerant object info and object-tolerant location info
- doesn’t matter where object is, can atill recognise
16
Q
what is the single cell representation of objects?
A
- Hubel & Wiesel found hierarchical processing in visual cortex
- simple cell to complex cell to hyper-complex cell (which has n-stopping property)
- pattern-processing in inferotemporal cortex
17
Q
what object features are the different regions of the temporal lob responsible for?
A
V1: edges
V2: contours
V4: colour & shape
PIT: simple features
AIT: elaborate features
18
Q
what is the cell selectivity for pattern processing in the inferotemporal cortex?
A
- cells respond to code shape, colour and texture
- respond to all objects with these properties
- generalise across position
- organised in columns through cortex surface (e.g columns of object-selective cells in AIT)
- posterior region cells more orientation- and size-specific (general), anterior region more responsive to specific objects
19
Q
what are the steps in Riesenhuber & Poggio’s 2002 hierarchical model of object recognition?
A
- input image, first processed by simple cells (different cells for different areas of visual space) => different orientations in different space coded for
- complex cells provide input into hypercomplex and V2 cells
NOTE: two options for simple to complex cells - MAX operation (complex cell will respond to which simple cell responds most)
- pooling (weighted sum of info, adding composites together)
20
Q
what are the properties of the hierarchical model of object recognition?
A
- anatomically and physiologically plausible- based on known connections and properties of brain cells (from V1 to IT)
- based on earlier hierarchical models
- copes with viewpoint dependence and independence
- incorporates theories of learning
- copes with multiple objects and objects in different context
21
Q
what is the role of context in object recognition?
A
- context can be normal or abnormal
- wihtin scene -> recognition of objects is easier in correct context
- within object-> word context biases interpretation
- word superiority effect- detecting a letter is easier when in a word
22
Q
explain bottom-up processing of letter recognition
A
- provided with stimulus, detected by low-level feature detectors that have features similar to letter
- these low-level detector cells connected to mid-level pattern detectors which have a nice representation (more like the letter)
- mid-level excites the correct high-level object detector and inhibits the incorrect ones
23
Q
explain top-down influence on letter recognition
A
- we have memorised concepts in temporal lobe (e.g what has a tail? rat and cat-yes, mat- no)
- excites possible highs, inhibits ones that don’t match concepts
- excites suitable mids
- anatomy- more connections descend than ascend
24
Q
what happens when we combine bottom-up processing and top-down influence in letter recognition?
A
- bidirectional processing models: info flow is bottom-up and top-down
- expectations (of what something might be) lower threshold for likely items (allows to detect better)
25
what is the role of temporal context and object permanence in expections?
- we expect an object moved behind a screen to reappear with the same form when screen removed
- Bower: >6 months olds surprised if object gone if screen removed (expects object to be there)
- indicates at this age, they have permanence of own toy (know objects exist)
26
how does similarity factor into object representation?
- similarity underlies how objects are represented and organised (objects are represented in brain in terms of similarity)
- therefore guiding classification, naming, behaviour
- similar objects = similar neural representation
27
what makes objects similar/dissimilar? explain the experiement used here
Cichy 2019- investigated link between brain acitvity and different object properties
- groups: asked to judge similarity of shape, function, colour, background or freely without instruction
- asked to put similar objects close to each other within circle depending on group condition (close together = similar)
- compared perceived similarity (free vs. other to see which closest)
- 'free' group similar to 'function' group
28
what do we find when we compare psychological similarity matrixes against similarity matrixes from different brain voxels?
- lower perceptual distance = more similar
- perceived similarity of objects related to ventral visual cortex activity
- representations emerge within 200ms- object colour first (earlier in hierarchy), then shape, then background and free-arrangement in ventral temporal cortex
29
how might AI perform object recognition and what do we need it to do?
- should provide us with accurate and meaningful information
- it understanding context/place can be used to help determine objects
- usual method is supervised learning- train network on known objects
30
what are the two requirements for supervised learning?
1. algorithm must be suited to task
2. must have training dataset with appropriate 'coverage' (led to rise of multi-million datasets of objects)
31
what is a convolutional neural network (CNN)?
- a deep-learning algorithm that learns features directly through data
- (for objects) CNN + 1.2 million object database + 2.5 million scene database ~ human performance
- take image and convolve into different layers that represent edges
32
what do deep neural networks (DNNs) do?
- predict behavioural assessment of perceived similarity
- show spatial invariance: can extract relevant features, handle noisy/cluttered images and image classification/object detection/semantic segmentation
33
what do you need in the AI to enable to carry out CNN/DNNs?
- large computers
- that do not 'over-fit' (trying to detect patterns when there aren't any)
- layers to be interpretable
- ability to handle irregular structures
