High Level Perception Flashcards
1
Q
Object recognition
A
- Object recognition: is the ability to know what an object is – it involves identifying the shape of the object (despite changes in sensory input) and retrieving information from long-term memory about the object (e.g., its function, size, colour etc.).
- Object recognition takes around 200ms
- Why is it important?
- Neuropsychology - Brain injury / damage – effecting object and face recognition
- Computational modelling / machine learning / robotics
- Feels effortless- easy to label objects
- Certain pattern of photoreceptors triggered in the eye and then long-term memories about certain objects are identified e.g. a chair
- Computationally difficult- hard to program a robot to recognize an object because the sensory input on the retina changes in different lighting and view angles
- Take object recognition for granted- actually very difficult even though it seems easy
- People who work in artificial visual systems know how difficult it is- in many ways a very applied topic in psychology
- Brain damage often affects object recognition e.g. object agnosia- cannot recognize objects after strokes affecting the posterior cerebral artery
2
Q
object constancy
A
- Object constancy: is the ability to recognise objects across variation in sensory input caused by changes in light (shadow), scale (size), viewpoint and occlusion.
- Essential for object / face recognition and high level vision
- Ultimately the visual system has to achieve object constancy- recognise objects despite changes in brightness, viewpoint, scale and occlusion
- One problem for the visual system is viewpoint- seen in different arrangements but can still be recognized and named
- Retinal images are very different but still class it as the same car/dog
Challenge is how we achieve visual constancy
3
Q
object recognition- the problem of shadow
A
- When there is shadow it is challenging to identify the edges that define the object shape
- The visual system must work to figure out which edges belong to the object and which belong to the shadow
- (compare the images on the left and right – the right one shows all of the edges, and some of them belong to the shadow outline).
- How does the brain know it is a shadow not a protrusion out of the object?
one way or another we are very good at processing shadows- most likely explanation is that they move along with the object
4
Q
object recognition- the problem of variation in scale
A
- object size changes on the retina depending on how far away they are
- Things that are further away are smaller on the retinal image compared to those that are close by
- Talk about the size of things using degrees of visual angle for a viewer at a certain distance
1cm= 1 degree of visual angle
5
Q
object recognition- variation in spatial location
A
Object recognition is also difficult because of variations in spatial location– object position on the retina changes as objects move about
6
Q
object recognition- problem of occlusion
A
- Object recognition: the problem of occlusion caused by scene clutter – foreground objects partially occlude background objects
- NOTE! We can still recognise objects despite incomplete sensory input!
- Occlusion is a big problem- a lot of objects are often partly occluded when you look at them
- Objects often self occlude e.g. the back of an object is occluded by the front
- Very good at coping with occlusion- somehow mentally fill in and estimate the shape of occluded regions
Challenge for visual science is how the brain achieves object constancy
7
Q
gestalt laws of perceptual organisation
A
- These are principles that describe how the visual system can use ‘bottom-up’ processes to group image features into shapes and forming a whole
- Related for educational purposes
- Gestalt psychologists were German psychologists who talked about how people do not see parts they see parts grouped into gestalts
- Interested in the way the visual system groups contours and vertices as belonging to certain objects
- Also interested in figure ground organisation- identifying certain things as the figure of a painting compared to the background
- Often hard to see what gestalt laws are
- Olympic games rings- natural to interpret this as 5 rings overlapping and touching each other. Other interpretations are also possible but you go for the simplest interpretation of it- unconsciously the brain applied a rule saying the simplest interpretation si the one to go for. Lots of perceptual interpretations of the retinal image but you always go for the simplest e.g. if you see something change shape on the retina you believe it is a rigid object from a new viewpoint not the idea the object is plastic and is morphing into some thing else.
8
Q
evaluation of gestalt laws
A
- Gestalt laws are more descriptive than explanatory- we can describe them and illustrate them but not say why we do that
- Sometimes the laws also conflict each other
Gestalt laws also tend to emphasise hard wired rules- did not focus on learning and attention e.g. underestimating the power to learn new rules which could be more important than the gestalt psychologists foresaw
9
Q
basic processes- perceptual organisation
A
- But perception of shape involves both ‘bottom-up’ and ‘top-down’ knowledge
- Involves top down knowledge as well as bottom up rules
Also flexibly interpret the scene which is not the same every time
10
Q
anatomical basis for high level vision
A
- Primary visual cortex at the back
- Dorsal stream leading from the occipital to parietal lobes (responsible for rapid action control- using visual information to guide where you move)
Ventral stream leading from the occipital to the inferior temporal lobes is responsible for conscious vision and object recognition e.g. faces and letters - LOC- know that this is activated in fMRI when you show people line drawings of objects compared to scrambled line drawings
- Sensitive to shape and structure in familiar objects
- Next to LOC is the fusiform face area- sensitive to faces
Parahippocampal place area- activated in spatial navigation tasks e.g. mapping your way to a location (Maguire and London taxi drivers)
11
Q
use of vertices in object recognition
A
- Not a familiar object but can still recognise its shape
- One crucial bit of information is the vertices (corners where straight lines touch)- we know they are important from different demonstrations e.g. attneaves cat
Low-level image features such as edges and vertices (intersections between edges) provide important shape information
12
Q
Attneave (1954)- sleeping cat
A
- A simple demonstration that we can derive a lot of useful information about shape from the edge information alone.
- “Line drawing of a sleeping cat can still be identified when the smoothly curved contours are replaced by straight-line segments”
- The crucial points of maximum curvature and make them into dots and draw straight lines between them- make an arrangement of lines that is identifiable as a sleeping cat
Shows vertices are important for recognizing objects
13
Q
Biederman (1987)
A
- Line drawings just as easily recognized as ones that are colored and shaded in
- The image of the right is easier to identify despite containing less edge contour . This shows that vertices are important, not just edges …
Edges and vertices are important sources of information for shape recognition
14
Q
shape can also be described using other kinds of features
A
Study note: These shape features are sometimes called ‘primitives’ which means elementary ‘units’ of shape. A volumetric primitive is one that has 3D volume (e.g., a volumetric part). E.g., we could say a table has five volumetric parts (four legs and a top).
15
Q
recognition by components: a theory of human image understanding (Biederman, 1987)
A
- His big idea is that if we can recognise certain prototypical 3D objects and how they are fitted together that is sufficient for recognising any possible global objects- innate volumetric shapes saved in our memory and once we know these we can recognise any object e.g. like an alphabet but for objects
- He called these parts geons (geometric ions)
- Geons have certain characteristics that remain variant despite distortions e.g. numbers of surfaces
You can think of geons as basic shapes like cylinders, bricks, pyramids etc., that you could use to ‘build’ more complex objects (like ‘Lego’ bricks!).
16
Q
Marr (1982)- structural description theory of high level vision
A
- Marr’s model is another structural description where complex 3D object shapes are represented by their parts
Marr had a similar idea- our representation of objects is ultimately a representation of a number of deformed cylinders and how they attach together
17
Q
face recognition
A
- Faces are not just any stimulus they are special as they activate social scripts and emotions and help us identify different people- the most important stimulus in your environment
- Much better at identifying and remembering faces than other objects
Faces are special stimuli- not made of geons in the same way potentially
18
Q
the problem of face recognition
A
- Face recognition presents a difficult computational problem for the visual system:
- Faces are visually similar – requires subordinate level classification.
- Perceptual input highly variable across viewpoint.
- Faces are dynamic, moving stimuli.
- Faces are variable because of emotional expressions and speech.
- Remember: Subordinate-level classification occurs when we have to distinguish one category member from another (e.g., the faces of two different people, or two different kinds of car). Basic-level classification would be recognition of a particular category (e.g., distinguishing a table from a chair). Face recognition is a subordinate-level classification problem. They also move and express emotion (unlike other objects – so faces pose a challenge for recognition.
- Very difficult to see the difference between faces
- With some expertise and training you can make finer level distinctions within the subordinate category
- Have to cope with highly variable viewpoints and angles
They also do not stay still
19
Q
an information processing model of face recognition (Bruce & Young, 1986)
A
- Occasionally name generation fails- someone is familiar but cannot remember their name
- Very often if you can identify their name you can remember lots of other information about them
- We can damage fusiform face area and no longer remember faces- very debilitating
- There are separate representations of expression and facial identity
- In this model face perception involves the abstraction of facial expression
This should remind you of the general vision problem of abstraction – to recognise a face do we have to ‘remove’ emotional expression? This is like with objects when we asked whether we had to abstract (remove) shadow and orientation from shape. Same old problem! Does your ‘mental representation’ of your best friend have an expression or is it ‘neutral’ (in some sense). What information about your friend’s face does it contain? Also note – you do NOT have to learn this model – we are only using it to illustrate how face identity and expression might be separately stored in memory.
20
Q
effects of face inversion
A
- Facial recognition works best when the faces are not inverted- dramatically reduces the ability to recognize faces
- Brains face representations have a certain orientation
- We find it harder to recognise inverted faces than inverted objects
- Face perception is ‘tuned’ to the upright orientation. We do not notice feature inversion in the upside down image.
This suggests that face processing relies on configural or ‘holistic’ information about the whole stimulus, rather than an ‘analytical’ processing of each face feature (part).
21
Q
the perceptual expertise hypothesis
A
- Based on these results the ‘Perceptual Expertise Hypothesis’ was proposed. It states that there is NO domain-specific processing for faces. Rather, the face-specific effects previously found in other studies solely reflect subordinate-level processing of faces (since subordinate-level processing of novel objects like Greebles gives you the same results!
Here is a definition of the ‘Perceptual Expertise Hypothesis’. In brief, it says that it is not FACES per se that matter, but that faces are processed at a subordinate-level. So anything processed to a subordinate level should give face-like effects (at least, according to the Perceptual Expertise Hypothesis).
