Viewpoint dependency debate/ Faces are special

• Object recognition is viewpoint invariant
• • viewpoint independence
  • never seen those specific pictures before but you can still identify the objects
  • e.g Bruce’s idea of the structural code in face recognition
• • AI and viewpoint invarience
  • Marr
  • origin of the viewpoint invariance view comes from AI
  • Marr believed that vision operated in an information-processing manner

  
• • A grey level description is an image depicting light intensity changes
  • Each point (pixel) on the image indicates how much light was received at that point
  • This is computer vision, so, unlike Bruce’s Pictorial Code in face recognition, it is a computer image made up of pixels
  • looking at greyscale to find the parts that stick out
• Raw Primal Sketch

  • Find important parts of the image
  • done by looking for large intensity changes
  • Description of the boundaries between objects
  • achieved by several types of blurring
  • blobs, edge segments and bars can be identified
• Proximity

  • These blobs, edge segments and bars need to be put together
  • Gestalt laws propose ways in which this might happen
  • report they see a square made of circles for A
  • but then saw they see 5 rows/columns of 5 circles for the other 2

  
• • square background made of squares and a square made of circles
  • X in front of a square
  • similar things go together

  
• • People will see the circle as closed on first glance

  
• • if it has the chance to carry on it will
  • could turn anyway
  • so not good continuation
  • people will see AOB and COD

  
• • most people are asymmetrical
  • those who have average symmetry faces are usually ranked more attractive

  
• Putting the bits together

  • Similarity
  • Closure
  • Good continuation
  • Symmetry
• The Full Primal Sketch

  • uses Gestalt principles to determine important regions or lines
  • contain 3 different levels of depth
  • closest, medium, furthest away
  • shown in separate images
  • 3 images on top of each other
  • 2.5D sketch
  • As the name implies, is somewhere between 2D and 3D
  • A 2D image is flat and contains no information about depth
  • A 3D image shows full depth
  • The 2.5D image contains information about the layout of object surfaces but is still viewpoint-dependent
  • That is, the appearance can still change with a shift in viewpoint

  
• • Object recognition
  • turning objects in a 2.5D sketch into 3D objects
  • Marr and Mishihara 1978
  • proposed they could use structures called primitives to active this
  • everything is made of cylinders

  
• • Recognition by components
  • Recognition by Components (RBC)
  • Developed by Biederman (1987)
  • Objects are also comprised of features
  • An ‘alphabet’ of 36 geons was proposed
  • All objects can be constructed from this set of geons
  • The geons are like to parts of a letters in a feature net
  • An object is recognised by the geons activating parts of the network
  • A key concept of RBC is viewpoint-independence
  • That is, objects can be recognised in any viewpoint
• Recognition by Components (RBC)

  • Developed by Biederman (1987)
  • Objects are also comprised of features
  • An ‘alphabet’ of 36 geons was proposed
  • All objects can be constructed from this set of geons
  • The geons are like to parts of a letters in a feature net
  • An object is recognised by the geons activating parts of the network
  • A key concept of RBC is viewpoint-independence
  • That is, objects can be recognised in any viewpoint

  
• • Biederman 1987
  • used a flashlight and a more complicated penguin geon
  • recognition getting worse the more geons you have in the shape

  
• • viewpoint invariance
  • In summary, work by Marr, Marr and Nishihara, and Biederman propose that object recognition is viewpoint-invariant
  • Objects can be successfully recognised from any viewpoint
  • There is no cost of changing viewpoint between study and test
  • This is achieved by a group of set shapes known as geons
• • viewpoint dependency
  • Object recognition is dependent upon having a stored mental representation of an object that is the same/similar to a currently presented view
  • Recognition to a previously unseen view requires generalisation
  • Already encountered one form of viewpoint-dependence in PSYC412
  • (Bruce’s idea of the pictorial code in face recognition)
• • Longmore et al 2008
  • rotating head
  • shows profile view then other rotations
  • gets worse the more you rotate away from the original image

  
• • Canonical Views
  • most people choose 3
  • canonical viewpoint
  • shows the most info in one image

  
• Problem for RBS

  • gets worse with the greater angle of difference
  • demonstrates viewpoint dependency
• Liu 1996

  • made less mistakes with more symmetry

  
• • Reaction times and recognition accuracy is a function of how similar the study and test images are.
  • There is a steady increase in reaction times and a decrease in recognition accuracy with increasing viewpoint changes
  • Suggests that object recognition is mediated by mental representations based upon view-dependent descriptors
• • The viewpoint-dependency debate was at its peak in the 1990s and has now largely subsided
  • It moved away from the idea of a view-dependent vs. view-invariant mental representation to what is the actual process used for generalisation
• Use both Dependent and INvariant processes

  • Burgund et al 2000
  • Image of objects tested in the same or different orientation as studied
  • images presented to the left or right visual field
  • processed by the right or left hemisphere respectively
• • mental rotation
  • Long before the viewpoint-dependency debate some element of viewpoint-dependency was known
  • Studies on mental rotation seemed to suggest this
  • Shepard and Meztler 1971
  • abstract objects (Rubiks snake) presented in pairs
  • could be same or different
  • angular difference between 2 objects could vary
  • Ps task to determine whether the ibjects are the same or different
  • the time participants took to judge whether two rotated abstract block figures were identical increased with the figures' relative angular disparity
• • Are there dedicated networks that are innate for recognising faces.
  • It is difficult to define faces as special as in face recognition faces are special because they convey important information(as stated by bruce and young)
  • special=dedicated neural networks dedicated to processing specific stimuli
  • Evidence for faces being special
  • Some evidence for faces being special is the N170. this is an ERP(event related potential). When people are shown different stimuli the brain processes the information on different levels. It also has negative potential meaning down is good.
• • FFA specificity:
  • Fusiform face area(fusiform gyrus) responds largely to faces. kanwisher(1997) believes this is because the FFA was designed to process faces while gauthier believes that it is a general area for expertise(anything you are good at)
  • Developmental evidence:fantz(1961)
  • Fantz was interested in what babies could actually see(because of this we know babies have blurry vision) fantz narrowed bars on a stimulus until the baby couldn't see the difference and no longer paid attention. He then decided to see how long babies looked at different stimuli for lengths of time
• black bar= under 3 months  white bar= over 3 months

  • This shows that babies looked at the face the most at any age, even though the top 3 stimuli all had the same contrast(black and white) the face was still on top, suggesting that they favoured the face.
• Johnson et al:1991

  24 infants less than an hour old

  • camera at the top records the babies reactions, the mother holds a paddle with either a face(or part of a face/shapes of a face) or nothing on and the mother would move the paddle over the baby to see if the baby reacted.
  • 0= the baby looking directly up
  • The face lead to the baby moving its head and eyes more than anything else. This suggests that babies are born with what faces look like and it is an innate process.

  This then leads to 2 systems being presented by Johnson and Morton(1991)- Conspec and Conlearn
• CONSPEC:

  • An innate subcortical system
  • It drives an infants attention to faces to aid with survival(drawn to other people)
  • It is necessary to drive the shaping of face processing(it understands face shape but are not able to recognise a face yet)
  • All this suggests that there are innate schematics for face processing
  • However, there is alot of critics
• CONLEARN:

  • Cortical, face specific mechanism that takes over from CONSPEC at 1-2 months
  • It is more flexible and therefore more accepted
  • When there are visual issues from birth, face learning is impaired as it is not able to act

  One of the problems with CONSPEC was shown in a study by simion et al(2003)

  • Infants were shown top heavy or bottom heavy
• stimuli.

  • It was found that infants preferred the top heavy pattern because there is the most contrast there and there is the most contrast at the ‘top part’ of the human face
  • indicating that it is not the face they are focusing on but the contrast
• Another issue with CONSPEC was shown by pascalis(1994),

  • they reported that 3-4 day old infants will habituate to a strangers face(they would ignore it),
  • their argument that CONSPEC is supposed to draw the infant to the face and CONLEARN is not developed yet so the infant cannot have learned the face.
  • But the infant was shown the face alot so it might not be interesting anymore
• Perceptual narrowing:

  • Perceptual narrowing(developed by nelson, 2001) has lots in common with CONLEARN.
  • This is the gradual learning of human faces, even if you don't recognise human faces, you can tell them apart.
  • As you start to develop you are better able to define the difference of different faces(pascalis et al 2002).
• Kelly et al(2005)

  • showed that 3 month old infants prefer faces of the area that they grow up race to faces of other races as that is the race that they see more often, but this preference was not present at birth.
  • Meaning the most commonly seen race is prefered by the infant.
• Neuropsychology:

  • Neuropsychology studies people with brain injuries and tests what they can do vs what they cant do and then compare that to the general public to see what damage does what.
• Visual agnosia = when presented with an object he is unable to say what the object is

  • Sheep recognition!!!!!!!
  • Sheep farmer who cant recognise his wife but he can recognise his sheep(in rather face like ways)
  • he knew they were different
  • WJ the sheep farmer
  • mcneill and warringon (1993)
  • WJ had a stroke and became a sheep farmer
  • recognise the identity of his sheep but not his friends/family.
  • Results:
  • he recognised his sheep and other sheep.
  • evidences the fact that there must be a face-specific impairment as face recognition is down to chance while sheep recognition is high
• Mr W prosopagnosia

  • Developed face recognition problems as a result of heart problems.
  • He had a number of abilities-
  • Normal IQ
  • Could identify faces from other objects
  • Could copy faces
  • Could state the sex of the face
  • Could state the expression of the face
  • Could match unfamiliar faces
  • He could recognise people from their voice(he hasn't forgotten about people)
  • However he could not recognise the faces(he couldn't identify himself or doctors or friends) therefore it is face recognition, not everything facial, just recognising them.
• Propopagnosia - inability to recognise faces