3D vision 1 Flashcards
what does a binocular observer gain?
binocular stereopsis (and convergence of eyes)
simple geometry would suggest that binocular stereopsis apparently provides complete information about the relative distances of P1 and P2 - as long as we know the distance between the two eyes
However BS is often said to only provide relative rather than absolute information
this is because we can converge our eyes such that fixation point falls on the fovea of both eyes
whatever our eyes converge on has zero disparity
non-zero disparities provide information relative to point zero
what is the vieth-muller circle
the locus of all positions in space that have zero disparity such that they stimulate anatomically corresponding points on the two retinae
zero disparity does not tell us that two points are the same distance
what are horizontal disparities
disparities measured around a horizontal axis on the two retini (which are separated horizontally)
disparities tell us about the position of a point with respect to the Vieth-Muller circle
outside V-M circle: uncrossed/divergent disparities
inside V-M circle: crossed or convergent disparities
do disparities tell us about depth
No - The same disparity is created by a small amount of depth at a close viewing distance or a large
amount of depth at a far viewing distance.
what is the inverse square law
the disparity of the retina is inversely proportional to the square of the viewing distance
do we use disparities?
to test must control stimuli and remove cues about depth and distance
Wheatstone’s stereoscope allows images to be presented to the left and right eyes independently
doesn’t show binocular disparity is sufficient for depth perception
need to remove other depth cues
do random-dot stereograms provide evidence we use disparities
pictorial stereograms contain other cues/sources of information
important to use random-dot stereograms
in a key region, dots in the left eye’s image are shifted relative to the dots in the right eyes image
this should produce the impression of a square floating above the background
when this is convergent: the square appears to protrude
when this is divergent: the square appears to recede
how can red-blue glasses create the same effect as random-dot stereograms
Essentially the same pattern but one is blue and black while the other is red and black
In the image presented to the left eye some of the dots have been shifted from the centre
The red dots got through the red filter and the blue dots go through the blue filter
When these are viewed as superimposed there is a central square floating in depth
Perceiving the depth distance rather than the shape of the object
what do RDS tell us about stereopsis
stereopsis can occur according to a new model where object recognition happens after binocular combination and stereopsis
this differs from the traditional model where it was though that monocular object recognition occured before binocular combination and stereopsis
Not true that random dot kinematograms have don’t have monocular features
There must be some features that are compared between the two eyes: dots, edges, micropatterns
Can’t achieve a process of monocular combination if there is no structure in the eyes able to do so
Arguing about a continuum of extent of monocular processing
What is being compared?
what is being matched in RDS
not individual dots as they can be seen over a wide range of dot sizes and densities and look identical
could be clusters as local clusters look very different, may be that the statistical properties are inferred rather than the shape
similarity of contrast is arguably essential
how well do we see depth in RDS
Test whether participant correctly reports whether centre square is in front or behind.
Manipulate the stimulus to make the depth discrimination easier or harder, and search for the point
where the participant gets it right 75% of time.
(note that 50% correct is chance performance).
features that are matched must be:
similar size
similar shape
same contrast in two eyes
what is the correspondence problem (Julesz)
for every dot in one eye’s view, there are thousands of possible matches with dots of the same brightness
what are the solutions to the correspondence problem
Julesz suggests two processes
local matching
global matching - (disambiguates local matches)
problem is ill-posed
problem of matching/ correlating individual dots
if, instead local areas are matched the problem can be solved (best local correlation)
what is Julesz’s shift and subtract model
based on his solution to the correspondence problem
the amount of shift needed to bring areas into alignment is equivalent to the disparity
what is Julesz’s spring-couples dipole model
local matches COOPERATE to achieve a global solution
what constraints is marr & poggio (1976) model of stereopsic matching based on
investigated constraints to exploit to find a solution that corresponds to properties of the world and our visual system:
similarity
only matching features which are similar
Eyes are closely spaced so images to the two eyes are generally similar
uniqueness
only match one feature with one other
we live in a world of opaque surfaces, therefore there can’t be more than one match down a given line of sight
smoothness
select neighbouring matches that are similar in disparity
depth generally varies smoothly because of the world of surfaces
what is marr & poggio (1976) model of stereopsic matching
incorportated these constraints into a biologically plausible model containing binocular cells or nodes of all possible disparity matches
similarity is incorporated because only white dots can be matched with white dots and vice versa
uniqueness is incorporated through inhibition down line-of-sight columns
smoothness is incorporated through excitation between cells responding to the same disparity
running all constraints produces black circles as most probable
Good model but limited to a single spatial scale as it only works for dots of a given size
what is Marr & Poggio (1979) new radically different model
separate processing in the left and right eyes
monocular spatial frequency channels at different spatial scales (Low and High)
local independent spatial frequency tuned disparity matching - separate matches at different spatial scales
unitary global broad-band stereopsis
stage 1: create multiple representations at different spatial scales (coarse, medium, fine)
stage 2: matching between zero crossings (edges) separately for the different spatial scales
matching is multiscale
order of processing is from coarse to fine scale matching
no correspondence problem as there aren’t multiple possible candidate matches
at the coarse scale there is one match between L and R
this coarse scale match then guides vergence eye movements to bring middle and finer scale representations in the two eyes into approximate alignment, restricts the range of disparity matches so we don’t have to consider all possible matches
disparity matches are limited to the size of the filter (size-disparity correlation)
false matches never arise - avoids rather than solves correspondence problem
summary
3D vision is not a special problem, need to consider if there is information available not whether the retina is 2D
binocular disparities proide information about location with respect to the vieth muller circle
RDSs tell us that some monocular processing must take place in stereopsis and similarity is important in the matching process
however the correspondence problem may not be a real problem that has to be solved explicitly depending on what you assume is being matched
