Object Recognition Flashcards
What ist he main challenge the visual system is faced with?
Quick identification of any object at any size, orientation, location etc
What are Gestalt’s laws of - like what are tehy for?
Perceptual organisation
What are Gestalt’s 5 main rules
Similarity Common fate - same direction Proximity Good continuation - same path Pragnanz - 'good form
What is the basis of the Pragnanz rule?
Koffka (1935) – of several possible geometrical organisations the one that willl actually occur posseses the best, simplest and most stable shape
Who came up with the new gestalt laws?
Palmer and Rock
What were the new gestalt laws?
Common region - same region of space
Connectedness
Synchrony - occur at same time
TWo limitations of Gestalt approach?
Doesn’t relaly explain how we perceive objects, more just a description of perceptual process and perceptual phonemona
Difficult to say sometimes what constitutes ‘simplest’/’best’ or what ‘good continutation’ is
What is the next advancement from Gestalt’s approach?
Triesman’s Feature Integration Theory
what is the order of events in Triesman’s view?
Object – identify primitives – combine primitives – perceive object – compare to memory – identify object if match
Where does attention come in or not come in in Triesman’s theory?
Identify primitives = pre-attnetive
Combining primitives = focused attention
What is Triesmans’ theory based on?
Feature ‘primitives’ – combine to make defined object
What are teh two main ways of determining primitives?
Image segmentation – 2 regions contianing different primitives - boundary will appear
Pop- out – if efficiecy of visual search = independetn of no of items presented = pop out
What are some examples of identified primtiives?
Curvature, tilt, colour, line crossings, line ends, movement
What can explain illusory conjuctions with regard to visual primitives?
Primtiveis = procesed independently
If stimulus possesses 2 primtivie features = sometimes combined inappropriately
Disadvantage of Triesman approach?
Largely tell us how we might combine features as opposed to how we recognise objects
Advantages of Triesman approach?
Basis of feature combining and role of attention in object recognittion
What are the two main models of object recognition
Bannerman Recognition by Component model
Marr’s computation approach
What does BAnnerman’s model involve>
It uses 36 simple geometric primitives to recognise complex objects
WHat are three poroperties of geons?
View invariance - can be recognised from almost all angles
Discriminability - can tell one from another from almost all angles
Resistant to visual noise
What are geons defined by/
Properties that dont change from view to view e.g. brick = 3 parallel edges, inner y vertex and 3 outer arrow vertices
How does Biederman’s model work? STeps?
Edge extraction – detection of non-accidental properties – parsing of regsions of concavitiy — determination of components – matching of components to object representation
What happens once an image has been parsed up?
You can detect the accidental propetires in each of the volumetric shapes to idetnify each geon. When geons can be extracted - subject can be recognised
What evdience is htere with regard to the importance of regions of concavities?
Biederman (1987) – removal of controus defining concavitities affects object recogntion.
No RoC = 50% recognition accuracy. Just RoC = 70%
Full pic = 100%
What can priming tell us about this model and who’s experiment showed us this?
Biederman & Cooper, 1991 – Pic for 500ms, 7 minutes, pic for 200ms. If pic 1 and pic 2 have same geons, then raction time will be faster when looking at 2nd pic. If they are sam object but with different geons, no priming takes place (semantic priminng does but no visual priming)
What is priming?
ACtivating a representation once makes it easier to use subsequently
3 limitations of Biederman approach?
different types of same object —> Goldstein 2010 – may bneed to be defined to show how to distinguish between faces = same features, look different
Neurons can distinguish between much smaller than neurons
No direct evidence for geons, only evidence of interpetation of this kind
What is the approach re computers?
Marr’s Computational Appraoch
How did MArr make his appraoch?
Based on how a computer would recognise objects – then compared to what humans are thought to do
What are the elements of Marr’s approach?
Image – primal sketch – 2.5d image/sketch – 3D model description – object catalogue
How do you get from image to primal sketch?
Find contours – best way = find second derivative - where rate of change of rate of change = zero. Zero crossings = simple cells detect them.
another way – where rate of change of brightness = maximum = edge = not economical way of finding them
How do you get from primal sketch to 2.5d image?
Detect these zero crossings at different spatial frequencies. When you combine them all = a number of images at different spatial scales.
Find regions where zero crossings form closed loops or where the zero crossings are the same across all scales
How do you get from 2.5D to 3D representation and what is the differnec ebteween them?
2.5d = viewer-centred. 3D = object centred
3D representation is developed based on a number of visual primitives derived from carving up the visual image
Who did an experiment on the formation of the 3D sketch?
Marr & Nishikara (1978)
What did Marr & Nishikara find?
Development of 3D sketch = based on processing of more elementary shaped primitives
Hierarchical organisation of primitives
Concavities are importatn in segmenting parts
Through splitting up images you get generalised cones not geons
3D model representaions = same regardless of viewpoint (object centred)
What does Recognition involve in Marr’s experiment?
Matching 3D model against a catalogue of stored 3D representations
What can help to distinguish between object classes in Marr’s view?
Lengths and arrangement of cones could distinguish between object classes
Two problems with Both Biederman and Marr’s models?
only really talking about basic level object recognition
Quite different to have a model to allow you to distinguish different exemplars in same calss
Who had a pretty pessimistic view on current object recognition models?
Tarr & Vuong (2002) “The truth is, at present, that no single model can explain”….
Was Marr’s work finished?
No, much of his work was unfinished when it was put together.
What is it called if we can recognise an object from it geons?
PRinciple of componentail recovery
What is the principle of coponential recovery?
Recognising an object from its geons
Who provided evidnce of objects nto being enocded as a whole?
Golcu & Gilbert (2009) –Novel objects that shared a component with the trained object were identified at much higher levels than those that did not.
Objects are not represented in holistic manner but it is the individual components that are encoded
Who concluded in support of not favouring either model?
MAyor (2010) - results suggests synthesis of elements of both of the 2 main theories rather than favouring one
What do we know about the role of colour in object recognition?
Bramao et al, 2011 –> Meta-analysis of 35 exps – Results strongly support the content taht colour plays a role in object recognition – suggesst that role of colour should be taken into account in models of visual object recognitiion.
What fMRI evdience do we have with regard to object recognitiion?
Joseph, 2001 – meta-analysis of fMRI studies. Recognition of manufactured objects activated more aspects of the ventral temporal cotex than natural objects or face recognition
What is the downside to Joseph’s evidence>?
The recognition task used was also a predictor for brain activity (viewing, matching, naming)
Conc – cognitive demands of a particular recognition tasks = as predicitve of activateion patterns as is category membership
