Question | Answer |
Difference between perception & recognition? | >Perception: what you see - shapes, colours etc (enough for you to navigate around the environment) >Recognition: what you see things as - dogs, faces etc (top-down, involves knowledge) |
Types of recognition - object & face | >Object: between-category; "what" - eg apple >Face: within-category; "who" - eg Freud Typically, separate research for each |
General recognition steps | >Basic sensory description (eg Marr's 2 1/2D) >Turned into 3D description/representation >Matches what's been seen before (IRRESPECTIVE of any angle) |
Recognising 2D objects - different cognitive processes than using 3D? | >Template matching: sensed image compared w/ range of templates until match found - unlikely as very generic templates or large # templates req'd >Feature recognition: key features extracted & compared w/ internal representation until match found - more generic; issue with ambiguity eg lines & curves >Structural description: structural description & key features w/ how organised in relation to each other compared w/ internal representation until match found - good re variety & ambiguity; good for 3D version of 2D object |
Marr's 2 1/2 Sketch - importance for recognition? | >Final stage of Marr's model of early visual processing - integration of outputs from various modules to form 2 1/2D sketch >Viewpoint specific >Enables interaction w/ environment (according to Marr) >STARTING point for models of object recognition - Marr & Biederman |
Problem with trying to recognise 3D object? | >Retinal image is 2D & objects look very different depending on viewpoint - primal sketch changes with viewpoint >Either we need to store: 3D representation of an object (viewpoint independent/invariant) OR many 2D representations (viewpoint dependent) >Marr & Biederman: agree 3D representation generated BUT disagree on process |
Marr (& Nishihara) Process? | >Object made of components = generalised cones 1) derive object's shape for 2 1/2D sketch. Assumptions made re points on the object which produces silhouette (Marr: contour generator) - eg each point on CG = different point on object 2) id the major axis/axes. (evidence: Warrington & Taylor '78; Humpreys & Riddoch '84) Areas of concavity link component cones/primitives Shape of object described in relation to axis/axes 3) Compare 3D structural description w/ mental catalogue to find match >Doesn't depend on viewing angle as description & model entries are 3D |
Evaluation of Marr (& Nishihara) | For: >location of central axis key to recognition >Lawson & Humphreys '96 - rotation of line drawings not affect recognition unless major axis titled toward observer; result of major axis appearing foreshortened? >Warrington & Taylor '78; Humphreys & Riddoch '84 - Ps w/ right hemisphere damage recognise typical but not atypical view of objects; similar for photos of objects Against: >Works well for animals, not good for furniture, fruit etc >Within-category discrimination - hard to explain conversion to generalised cones mapping to all exemplars of category to same representation Suggests we cant tell difference between 1 instance of a thing and another (eg all Westies are the same?) |
Biederman process? | >Same as M(&N) >Complex objects represented as hierarchies of simpler shapes >Concavity used to sub-divide objects >Recognition via comparing 3D representation w/ previous exemplars >Different to M(&N) >Geons not generalised cones >5 invariant non-accidental properties - no axis req'd >Clues from lines/outlines >Matching geons assembled into 3D representation |
Evaluation of Biederman | For: >Explains why objects harder to recognise when parts of image w/ greater concavity removed (Biederman) >Biederman & Gerhardstein Object priming - only works if viewpoint <135 degrees apart Performance decline if geons hidden between views Against: >Bulthoff & Edelman - Ps unable to recognise objects from novel POV >Tarr - Recognition may not rely on forming object-centred model; some factors viewpoint dependent (eg faces?) >Within-categorisation discrimination hard per M(&N) ie allows id of dog or cat but not 'our' dog or cat >Bottom-up focus=de-emphasis on import of top-down processes >Theory states objects consist of invariant geons - but object recognition more flexible w/ some objects not having identifiable geons |
3 stages of object recognition? | 1) Structural - id object as familiar 2) Semantic - access semantic knowledge about object 3) Naming - access object's name >Evidence from CNP studies - patients with breaks @ different stages |
Similarities of Marr (&Nishihara) and Biederman | >Similarity in approach >Info processing approach >Marr is more broad (covering all stages) - Biederman developed later stages of Marr's theory >Both assume shape is critical w/ starting point being outline >Both see representations being built from component parts - not holistic >Both deal with invariance via generation of 3D representation - ie not storing lots of 2D views >Both assume concavity important >Both assume recognition via matching 3D representation to representations in LTM |
Differences between Marr (&Nishihara) and Biederman | >How cues used to generate 3D representation - contour generator vs non-accidental properties >basic components (primitives) of 3D representation - gereralised cones vs geons >overall 3D representation - Marr = id of central axes w/ the representation being a description of components relative to central axis Biederman = geons & how they fit together |
Viewpoint-Dependent Challenge - Tarr & Bultoff | >Viewpoint invariant models cant explain all object recognition evidence >Tarr & Bultoff '95 - viewpoint-dependent representations req'd where object representations involve collections of views from specific viewpoints >Tarr - RTs & errors for naming familiar objects from unfamiliar viewpoint increase systematically w/ increased rotation distance from nearest familiar viewpoint >T&B - response to Biederman & Gerhardstein arguing evidence for viewpoint-dependence is strong Also - demos of invariance only work due to small exemplar set w/ few key features OR use common everyday objects where previous experience comes into play >T&B problems with Biederman: Geon structured descriptions (GSD) not sensitive enough - cant distinguish cow & horse >GSDs cant distinguish @ subordinate level - eg makes of car; faces |
Reasons for perceiving faces other than recognition? | >Expression analysis: determine people's emotions CNP - some lose ability to recognise facial expressions; BUT not same patients who cant recognise faces - suggests different systems >Lip reading: McGurk Effect Some patients don't show it & cant lip read - suggests different system |
Face recognition - 3 stages? | 1) Structural - id face as familiar 2) Semantic - access semantic knowledge about person 3) Naming - access person's name >Evidence - Diary study (Young) >CNP - patients w/ problems @ different stages >Experiments - naming slower than category judgements which are slower than familiarity judgements |
Familiar vs Unfamiliar faces? | >Good pictorial memory - Shepherd '67 - Ps 98.55 correct in force choice recognition of 600 pictures >Unfamiliar faces ≠ good >Burton & Bruce '99 - Watch video, then try to id 10 still pics. Mis-id 20% of cases >Kemp et al '97 - cashiers poor @ matching faces to photographs |
Problems with unfamiliar faces? | >If shown photo of someone we've just seen but from different angle, lighting etc then difficult to id >When only viewed once, a face will be represented pictorially rather than structurally - SO recognition dominated by external features & viewing conditions |
Bruce & Young model? | >Bruce & Young 1986 >Conceptual model (not computer) >Separate sub-systems for: *facial speech analysis *expression analysis *directed visual processing analysis >When face well known, it's encoded structurally. Recognition less dependent on viewing conditions |
Burton et al IAC model | >Based on Burton & Bruce 1986 >Connectionist model - interactive activation & competition network >Info flows top down >All about node activation - have to be active enough. If not then issues such as prospagnosia >FRU - face recognition unit >PIN - personal identifier node - info about them >SIU - semantic information unit - types of info eg names of occupation >WRU - word recognition unit (like FRU for words) >NRU - name recognition unit - linked to PIN >3 pathways >Face - FRU> PIN> SIU> Lexical output >Name - WRU> NRU> PIN> SIU> Lexical output >Other info - WRU> SIU> Lexical output |
Covert face recognition | >Recognition based on what someone tells us - ie conscious awareness >Recognition w/o awareness? >Prospagnosics cant consciously regonise faces - but they show different physiological responses to pictures of people they know - implicit/covert recognition >Modelled in Burton et al model |
3 questions re whether face recognition separate from object recognition... | 1) Special area of brain involved? 2) Ability learned or innate? 3) Do specific features help face recognition or do we recognise "whole face"? |
Are faces special Q1 Special area of the brain? | >Prosopagnosia = double dissociation; suggests different neural pathways >Capgras delusion - separate for objects & faces >Monkey temporal lobes - cells respond differently to monkeys & humans >fMRI - FFA activation |
Are faces special Q2 Learned or innate? | >Johnson & Morton ('91) - >newborns notice faces more than other stimuli >So, look @ faces a lot >So, learn more about expressions etc. >Different skill to other recognition - ie generic features @ 1 level (2 eyes, nose etc) >Yin ('69) - better memory for faces than other objects; inverting faces has more impact on recognition than inverting object - BUT different types of categorisation involved! >Diamond & Carey ('86) - dog experts disproportionately affected by inversion (care - results not replicated!) >Configural/holistic processing used by experts? We're all experts @ face recognition >Thompson ('80) - Thatcher effect |
Are faces special Q3 use of specific features or whole face? | >Yin ('69) - inverted faces delay suggests processing upright faces holistically >Farah ('93) - features better recognised in context of the whole face then when presented alone >Bruce ('94) - moving features up/down impacts recognition >All suggest holistic/configural processing |
Bruce & Young model (1986) - 8 components? | 1) Structural encoding - produces various representations of faces 2) Expression analysis - other people's emotional states inferred from facial expression 3) Facial speech analysis - speech perception aided by watching lip movements 4) Directed visual processing - specific facial info processed selectively 5) Face recognition nodes - contain structural info about known faces 6) Person identity nodes - provide info about these individuals (eg occupation) 7) Name generation - person's name stored separately 8) Cognitive system - contains additional info (eg most actors/actresses have attractive faces); influences which other components receive attention |
Predictions from Bruce & Young 1986 model? | 1) Major differences processing familiar & unfamiliar faces - unfamiliar requires more processing 2) Separate processing routes for processing facial id & processing facial expression 3) When looking @ familiar face, familiarity info from the FRU accessed 1st, then info about that person (eg occupation) from PIN, then person's name from name generation component Thus - familiarity decisions faster than decisions based on PINs, in turn faster than decisions re person's name |
Limitations of Bruce & Young (1986) model? | 1) Omits 1st stage of processing - ie when observers detect that they're looking at a face (Duchaine & Nakayama '06) 2) Assumption re facial id & facial expression involving separate routes too extreme? (Calder & Young '05). Majority of prosopagnosics have problems processing expression as well as id - 2 processing routes probably only partially separate 3) Assumption that processing names is always after processing other personal info is too rigid (Bredart et al '05). More flexibility req'd - proved in various models (eg Burton, Bruce & Hancock '99) |
Want to create your own Flashcards for free with GoConqr? Learn more.