Object and face recognition Flashcards
How does perception proceed to recognition?
- We recognise an object by comparing what we perceive with an internal representation of the object.
o Must have some sort of stores representation of things to recognise them - Humans can recognize the same object from many different perceptual representations.
What 6 issues are encountered in perception –> recognition?
Variation in lighting Variation in background Occlusion Variation in viewpoint Variation within categories Living things change slowly and quickly
How does variation in lighting cause problems?
o Lighting condition matters a lot
o Doesn’t change recognition but colour information changes
o So, we must sue colour in some way to recognise things
o But not completely reliant on colour as we still recognise them
o So, we must have a very robust system that doesn’t rely overly heavily on colour perception
How does variation in background cause problems?
o Issues in recognition against background – not always the same background
o The more complex a background is, the more issues there are with recognition
E.g. where’s wally
How does variation in occlusion cause problems?
o How the brain fills in the gaps
o When walking in the real world, we can still recognise objects even with parts obscured
o But some things may be missing key parts inhibiting recognition
How does variation in viewpoint cause problems?
o The angle doesn’t matter as long as you know what you’re looking at e.g. a teapot
o You can recognise your pet or an animal you know from any able because you’re familiar with it, but with ones you don’t know, it becomes more difficult
E.g. pet cat
o There is a big difference between familiar faces and unfamiliar faces
How does variation within categories cause problems?
o How do we extrapolate from our brains to new instances of things we see?
o E.g. stored representations of things must be robust to be able to recognise unfamiliar things when recognised as belonging to a category
o This is how me must store thing but also have other representations to be able to match unfamiliar things – can’t store all chars ever seen so we have to categorised it
How does living things changing slowly and quickly cause problems?
o Can recognise school friends years on in life, even when not friends
o You can still recognise adults even though they have changed from children
o Can also recognise people when smiling even when muscle movements in face change e.g. when smiling
o Brains can ignore the changes and still extrapolate the identities
What are the different models/explanations for pattern recognition?
- Template matching
- Prototype matching
- Feature analysis
- Recognition by component
- Gestalt principles
None of these are completely right or wrong, it is probably a combination of a few of these)
What is template matching?
- We store all copies of everything seen in brain and then compare anything new we see to these copies to recognise it
- The incoming sensory information is compared directly to copies (templates) stored in memory.
o However, huge assumption that we can remember everything – memory isn’t infinite and can’t store everything
o This does not allow for variation (e.g. in birds) unless there are templates for each variation (which is implausible).
o If it had to perfectly relate to a copy of something, we would never recognize anything new - We may store some copies and refer back to these
What is prototype matching?
- It does not require a perfect copy of the object, but a prototype to compare to for recognition.
o Prototypes are updated, changed and developed depending on what we see
o Prototypes may not be very distinct so that novel things can be easier recognised - If it has feathers, a beak, two wings, and can fly it is a bird. A crow is a more typical exemplar of the category ‘bird’ than a penguin.
- However, it does not explain
- Variation away from a prototype – important in recognising individuals in a category e.g. recognising a penguin, classified as a bird but also very distinct
o Typical exemplars and non-typical exemplars
o The further away you are from the prototype, the more distinct and out of the category it is
What is feature analysis?
- An object is deconstructed into features, some of which are more important than others for recognition.
- People can recognise things from basic features, we combine them into a whole concept allowing us to recognise it
o E.g. recognise beaks, wings, feathers
Recognise it is a bird - However, it might be difficult to come up with a unique list of features that can capture all the different versions of a bird.
o Issue in identifying unique things to separate specific things e.g. sharks and dolphins
What is recognition by components?
- Objects are composed by a set of volumetric primitives (Geons), which have unique collections of viewpoint invariant properties.
- Object recognition is impaired when removing viewpoint invariants such as vertices from the objects (Biederman, 1987).
- Some objects are critical, so adding the handle makes the cylinder look like a mug, and the square look like a briefcase
- Some aspects of images are more important than others
o Removing viewpoint invariant negatively affects object recognition. - By breaking the joins between types of objects, it influences recognition – when you move the features apart, things aren’t recognised. But when the features are put back together, they are recognised as an object
o Has something to do with the components of objects in helping recognition
What is the difficulty with recognition by components?
- It is difficult to use this approach to recognize different exemplars within a category (i.e., distinguish two birds or a coffee mug from a beer mug). Experience and expectations can help.
- Difficult to recognise different exemplars within the category
o E.g. if using only feature analysis, you can’t recognise if it is a beer or coffee
o So, has been argued that it is an initial pert of recognition which influences us in choosing what category we ‘look-up’
o E.g. Gestalt principles
What is haptic recognition?
- We can explore an object by touching it following stereotypical ‘exploratory procedures’.
- Partially sighted/blind individuals – important
o E.g. touching faces
Visual sense – when removed, you have to work harder to recognise things
What is auditory recognition?
- E.g. picking up the phone to a family member
- A high pitch and bright sound conveys the impression that the sound is produced by a smaller ball (Grassi, 2005).
- Increased loudness is attributed to bigger balls (Petrini, 2014).
- However, this isn’t as useful as visual recognition – sometimes pick up the phone to people and don’t recognise them straight away even though you may know them well
o This is less common when talking in person or over skype because you can see them - Another example of connection between sound and recognition
o The sound of something can lead recognising the size of objects
E.g. if a sound is louder or deeper, you may think the person is larger
How does recognition occur from biological motion?
- We can recognise things from the way they move, rather than shape information
o E.g. friends
o Point-light walkers & motion capture
Removes the person by making them points and watching them move (look like sick figures) - Observers can identify friends, recognize the gender, the actions carried out, and the social interaction between others (e.g., Cutting and Kozlowski, 1977; Dittrich, 1993; Pollick et al., 2005; Petrini et al. 2014).
How do we ‘see’ faces in everything?
- Recognise facial features even when things may relate very little to actual faces
- Central image off what a face looks like, and then we process other things that look a little like that which means we extrapolate this to objects like above to recognise it as a face
What is the inversion effect?
- Eyes and mouth are reversed – only thing changed is orientation
o Rarely see faces upside down so it doesn’t activate the same apart of the brain - Yin’s interpretation = when faces are upright, they are processed by special mechanism in the right hemisphere
o Much worse at facial processing when they are upside down
o Research to show that they aren’t even recognized as faces and are processed as objects - Faces presented upside down do not stimulate this mechanism and are treated like objects
What are humans good at and bad at recognising with faces?
- Humans are really good at recognizing familiar faces, despite changes in angle of view, light, and in the person’s appearance (e.g. age, hair cut etc…)
- However, humans are really poor at recognizing unfamiliar faces, and computer image analysis systems are even poorer (Hancock, Bruce and Burton, 2000).
o Even in a comparison task it’s hard to match unfamiliar faces
o Especially is one image is smiling and the other is not
What was found to do with facial recognition in a comparison study?
o People don’t tend to recognize a face by matching it to one out of a line up in everyday life
Apart from police line-ups. And a lot of this work is done on EWTs and line ups because these recognitions may be inaccurate
o The task is made to be deliberately hard
Even though they are different, they are chosen based on similarity (age, features, gender)
This doesn’t happen in real life, not all people share these similar characteristics
Could cause an issue with line ups – harder to distinguish due to similar features
What are internal and external features?
- Difference in what features we use for recognition when looking at familiar or unfamiliar faces
- Unfamiliar faces – focus on external features more
o Hair, beard, glasses - Internal features used when recognising friends
- Process of recognition (model)
What is the model of face processing?
Starts with structural encoding: “this is a face”. This splits into two:
1) Expression, facial speech, age, gender
2) Recognition
Recognition splits into the following chain:
- face recognition units: stored faces
- Person identity nodes: stored semantic information
- Name generation
Outline holistic processing with faces.
- Tanaka and Farah (1993) asked participants to learn faces.
- They then tested the recall of individual features in normal and scrambled faces.
o The location had an important effect upon performance.
o This effect disappeared when faces were inverted and when images represented houses.
o When upside down we do more component processing but when the right way up we do more holistic processing - Features are recognised better if they are presented within a whole face than if presented in isolation or within a scrambled face (Tanaka and Farah 1993).
What is the composite face effect?
- Upright faces are processed in an integrated “holistic” way, that prevents easy access to their constituent features.
o When the two halves of different faces are matched up, it is harder to see who each part belongs to, but when separated can recognise
How must faces be processed?
- Can’t be doing something as simple as feature matching to recognise faces
- When the face is inverted, reaction time in recognition was slower
- When stretched or thinned, reaction time isn’t altered as much
- The storage we have is something different to just having a snapshot to compare it to in the brain, it’s more flexible than that
How does colour influence facial recognition?
- Changing the colour of faces causes huge issues
o When making it negative, we can’t process it in the same way
o If face colouration is too far away from common colour for a human being, it means we can’t recognise them
What are people sensitive to with facial recognition?
- Haig (1986): people are sensitive to the precise location of the facial features, especially the eyes and mouth.
- Negative faces are poorly recognised - representation of shape from shading is important for recognition.
What has been found through the use of computerised prototypes of faces?
Computerised prototypes of faces
- Very generic, average looking face
- Could be a held average of a face which influences this
- As averageness is increased, the image/face becomes more pleasing
What does averaging of (computerised) faces do?
o We are quicker to rightly name a familiar face when average image than an original photo
o Recognition accuracy increases with the number of photos used to create the average image.
o Using averaged (over individual componenets) faces raised machine recognition accuracy to 100% (Jenkins & Burton 2008)
Removing variation makes it easier to recognise
What does adaptation to faces do?
- Large literature on exposure biasing subsequent perception in faces
- Adaptation of faces
o Adapting to different faces, influences subsequent perception
What is an ambiguous face perceived as when exposed to either a male or female face prior to ambiguous exposure?
- Ambiguous face
o When preceded by a male face, you perceive it as a female
o When preceded by a female face, you see it as a man
What is simplified face space?
- Axis shows the opposite to the face on either end
- Original and anti-face
o If you adapt to the anti-face you attend to the opposite face which leads to the perceptual recognition of the original face
What is the issue with simplified face space?
o Learning the faces (not personally familiar) might not be the same as personally familiar
Different perception type
o Stimuli – the faces are not real – can this impact on the results?
What is facial processing like in infants?
- Infants have quite blurry vision
- But they prefer to attend to faces
- They prefer to look at images that look like faces rather than scrambled, upright as opposed to inverted faces and prefer to look at their mum over strangers
What is the familiarity effect?
- People tend to prefer original versions of the mirror flips
- When exposed to anti-ben, people like anti-ben. But when exposed to real-ben, they like real-ben
- Can alter what people like in images and even what they find attractive