thank you mister chairman for the introduction
uh my name is because the company and i'm just assuming that it's and university
and this is a beautiful campus assumes both so you can um
well i wanna thank you for being here at the last the presentations and i'm
going to talk about the sampling pattern you i'm going to introduce this model in
a brief and clear way i hope to be useful at christmas from you and
this has been a work together we um but clean vocals and what ensure strong
well
first i would like to thank our funders which supported us to all this work
and this will be the conference i'm going to talk about first i will give
a little bit about the background the motivation of the work and um then i
will introduce the model
the spc model reviewing the school and the definitions and how we generate this pc
applied to some known cases and we will see if it is good enough for
us to extract features of interest and then i will compute the work together with
some hints of the future work
i'd like to mention that a group in a mixture then university's name to realistic
three D and we are more or less interested in a three D information movies
images or video and we kind of have the whole chain of a three D
starting from the capturing processing transmission and the post processing and adaptation to the display
and the viewing experience so we cover the whole chain but this will be about
the capturing where and when we talk about capture we mostly mean cameras and parameters
related to them so we will talk about cameras in this presentation and um when
we say camera there that can display now configurations about them you're familiar with some
of them porsche that all of the maybe the snr cameras or very available we
know about them but for specific applications such as really capturing or different kind of
capturing is we are going usually to um unconventional camera setups and it's important to
be able to model them you know the level that the light and to be
able to extract parameters with low complexity at the same time would level of description
and for example here this setup is uh used in and is it is upon
and um well they use this setup which is a camera here and the lens
every yeah to capture a digital photography and this is a very famous camera array
setup maybe we have signal before and this is a different camera setup you wanna
should diversity and the famous like true or maybe it's assumed to be paying the
slideshow camera which is a plenoptic camera so there are different configurations and uh things
about them and we would like to have a model to be able to extract
parameters
like what we have here um i want to say that there are different parameters
related to one camera and um i haven't seen that kind of map or kind
of measure to be able to put them in a scale and to be able
to compare demand for example say camera one is better than camera to in that
sense and if you do this to the camera and then these parameters are changed
in this way and that kind of behavioural at the same time a descriptive information
about the camera system and what is a camera i mean the camera system can
do and how is this so different setups for the camera and these are usually
parameters of interest for different applications and which i have pointed out here are more
or less related to the focal properties of the camera and i will come to
this point later for example for uh an application that angular resolution in a certain
plane is more important for us maybe we can extract parameters using the model and
see that camera to which is um shown by right column here is better for
this application at this is space from the object or
not able anyhow to compare camera one and two and come to a conclusion which
one to choose for what much modification to apply to the camera to get a
better result so maybe remarks about that work we are doing it is to keep
the complexity of the model low at the same time to give it a high
descriptive network which um can be used for extracting features or modeling the system
yeah that is widely used and uh i've seen many to be instance here is
for example there may be more models but these are too difficult examples one of
the ray-based model which is um considering light as the light rays for names and
the familiar with the two plane representation and having one point in each plane and
the line connecting these two points are considered as the right and we call with
this description and the method is usually approximated which we consider the light and on
the right angle to the optical axis is a small enough to apply some uh
approximations and it is widely used in different applications such as ray tracing photography microscopy
or telescopes and um they are all familiar with this model a more comprehensive model
more complex model is the wave optics model which is the using the light and
electromagnetic wave and the method which is working with this electromagnetic waves or
usually starting from maxwell equations and harmonic waves and fourier theory and it is able
to uh explain more properties well at the expense of more complexity and all
we are going to
somehow interim something between these two models and the scope of the work we are
uh doing is well oh
only in it is that it will be a geometry based model and it will
exclude anyhow the wave optics at least at the se stage and it applies to
the optical capturing system which can be as i said conventional cameras or new setups
so the motivation of the work is to have a framework for modeling complex capturing
systems and we uh expect that this model providers kind of two words to be
able to extract properties from the system and at the same time can be keeping
in mind low complexity and the highest level of the model
so basically what the model can be applied to different camera setups and we generate
the spc using tools or mathematics for geometry we have and well i was trying
to show something like the spc model which is the sampling pattern cube so and
put a Q and put this might samples inside which are in the form of
the light containers which are introduced to and from this model we are extracting features
i
well this model is helpful for visualising purposes also and also describing the sampling behavior
of the system
there can be wide applications for these models first of all study and design of
the capturing system can you want application and uh investigating system variations if we have
a system and the at some parts with or with very the distances are properties
of the system how it is reflected the sampling behavior of the system i
the one i pointed out at the second or third the slide it is investigated
inter camera records which one is better in what sense for this application or you
have to compare it to different camera setups and one possible application can be adaptation
of the post processing algorithms on which i will give some more about
well in this sampling pattern cube i'm talking about it is uh there is a
very simple idea and this is um originating from the light samples in this model
light samples in this model are in the form in the form of light containers
and uh we can put it along like tracing the rain model and what is
special about like containers is that they are focus light there are formal focused like
so there is a point which all by phrase or one of the light rays
are passing to read point and we call this point at the position of the
light container and there is an angular span associated with the lights container and in
this representation we have four angles associated with it but this is a representation at
this estate so a light container which we usually um once the bases in the
slice coming next as the tip position and angular span and information is somehow the
vocal properties of the system are somehow coded in this all samples and
well the light
containers will then produce the sampling pattern cube which i show but you hear and
is a small like containers inside distributed inside the um Q and so we can
say that the sampling pattern cube is a set of this slide container and i
will show how to generate again how to use it so basically we all we
had a camera and there is the space between the space in front of the
camera and we tried to um provide information about how this space example by this
camera using a light container elements
and for ease of illustration we
we have some simplifications here in that presentations i will the uh slice coming i
will consider only one row of the on image sensor oh to be a starting
weight and i will not go to the to an image sensor it would be
too complicated to put it on a plane to show that i won't show like
containers into the instead of three D representation by only go for xt representation and
two angles the starting and finishing and the chip so this is a simplification we
do for illustration purposes and
there is one more thing uh in judea space instead of X Z if you
have like containers like this predicted position and an angle starting to finish the angle
we
transforming them to position angle representation and disposition angle representation is basically x-axis and the
taxes here so that the X one like here the key position same we have
a span we have an angular span in that uh access and
instead of seeing cones like this we will have
piece of lines like this and
we should have in mind that is lines are
horizontal and be presented means the like is in focus it means there is one
positional information associated to the whole line which is the cheap of the light depending
on
so we will face only positions like this horizontal lines in the sampling pattern cube
and this one shows the simple idea behind how we generate the sampling pattern cube
we basically start from the camera if we consider this part inside the camera and
is the optical elements in the camera there can be only in main lens or
a combination of different lens setups and this is the sensor plane
we tried what we are going to do is to a form a light containers
on the sensor plane based on physical properties of the sensibly light acceptance angle of
the it's sensor so light acceptance angle and we define the first set up
like containers then we backtracked this like containers into the C and you see my
delight container is passing to an optical element that container transforms to a new container
for example this one is transformed to this one so in new T position and
angular information is associate that is um you meant to the light container and finally
in an iterative process we a project all the initial light contenders to the three
D scene in front of the camera and what we get is called the sampling
pattern cube we will work with that later well this is a more formal presentation
of the same process we have the flowchart we um
actually form the like and painters and then go to this iterative no and process
for a project all the like containers to the scene and finally we come up
with the set of light containers in the form of the specimen
well i will not go to the very detailed what to just give you some
idea we have optical um elements like lenses or like apertures and so on you
can refer to the paper for way more information but anyway for example here if
like containers comes to an aperture for an aperture you know where the aperture you
so we know which the plane the aperture temperatures located on and we know this
time
or
the opening gary of the aperture and the lack container coming to this aperture well
marcus cut out because it's not be in this kind of the aperture and part
is staying here so we will have a new one that container like this cutting
part which is not on inside aperture span and
we will come to see point which is the new like container and we will
go to the next iterated steps and for example for a lens
if this is the lens plane and we know the focal properties of the lens
and we know the lens equation then not contain upcoming this plane well you transform
to a new one
and you position and angular span is given to the new light container and we
go to this process
until all like containers are processed and no this is a very simple example the
schematic them out
a single lens system
if this is the image sensor and this is the single lens system
we have project information from the image plane to the three D space in front
of the camera or here if Z welcomes us consider the plane of the main
lens
and see what the minus the use the plane of the image sensor then is
lines showing the in focus light which is the form of the light container as
it so before are i projected to another plane so you see that the angular
span of the light that there is a big change as well as their positional
information has been changed and now we have a new set of light containers in
the form of the spc that we can extract properties of interest from
well um we want to show that
but like standing there is actually reflecting the behaviour of the system or in the
better weren't the spc model in general is reflecting the behavior the sampling behavior of
the system and to show that we are applying this explicit model to a known
cases plenoptic camera
in the uh can um conventional form and the focused plenoptic camera
i hope you're familiar with system setups um i give some method with details about
them but all well i think these are well known systems are system is uh
containing the main lens lenslet carry an image sensor is placed behind the main the
lenslet i and C two systems both have the same optical elements and the only
difference between them is that distances between the a colour elements here we have
uh it's between the lenslet area of the image sensor as the for the focal
length of the lenslets
here it's not the same we have space E and it is smaller than the
and there is a relay system relation between the image plane which is here
and the sensor image sensor
and the main lens is pushed with forward so the spacings and basically different although
the optical elements are the same and this is a slight difference gives them very
different properties in terms of sampling and
um high-level properties of the camera like resolution like um depth of field the local
properties
well these are just a bit more information about uh one of the camera
i for the first split up to cover the conventional form and i would like
to highlight that the spatial resolution is equal to the number of lenslets in this
setup after we render images the spatial resolution of the images are equal to the
um number of lenslets and there is a trade-off so if you raise the number
of lenslets the spatial resolution scoring higher but angular resolution is going to be over
so this is the main feature associated with pc i and i would come back
to this point later
and pc F structure
which as a set has a relay system between the main lens um we commend
as there is a relay system and um it can be considered as an aerial
cameras inside the camera so the behaviour is more or less similar to the camera
panning and there are multiple positional information for each angular information and the spatial resolution
is which is a decoupled from the number of lenslets in this setup is the
main difference between the two camera by and
there also is that the numbers you have used for our simulations these are typical
numbers and um there have been practical setups with this number of the basic thing
i want to highlight here is that the only difference between none of the camera
i and are in the spacings
between the main lens and then on the other hand from the lens the area
to the sensor and the rest of parameters are the same so what accuracies result
from the difference in the
spacings and these are typical
in spc shapes we expect from the camera i and a lot of the camera
that you see that we have a kind of area here sample and here we
have very narrow area in the form of a line may be spread in the
space
and uh the angular span we see here is where considerable and angular span of
the samples are very small here and oh
here is a closer look the same information we have this for plenoptic camera i
and we can see in there
instead
this is the area sampled actually um the density is too high so it's just
we have the shape here as a color but if you look from the inside
uh in setup is um
information we can see the sample like samples your in the form of the light
containers and we can see multiple angular samples for a single position is a single
position this is expenses so there are multiple angular samples forcing the position and this
one is showing samples coming from behind one lenslet so the information behind one lenslet
captured on the image sensor are the formal column
in the case of plenoptic camera i
and this is that we also you
basically the same data and this is the case um plenoptic camera have
and we see the sampling properties are different these are the multiple position samples for
one single angular span and we can see the sample data by pixels behind one
lenslet
i hope to give you the impression that um
spc is following the behaviour of the camera system so on the next slide is
showing actually if we apply variations in the camera system is variations are reflected in
the spc and this variation i've decided to be the variation of the lenslet pitch
size in this case we can see the information we bill in its size of
um interest haven't changes when the its size of the lenslets are battery and we
can see the trade off between angular and spatial resolution in one of the camera
i case lighting or not the camera is there is no trade off and um
it is confirming that this spc model is falling the behaviour of the system
but i do not talk about uh about the feature extractors are which is an
ongoing work also we are more or less now focusing on the resolution parameters and
um
this features extractors as a set can be informal focal plane field-of-view spatial resolution in
different depth planes and angular resolution and that solution and different focal properties and um
we hope to publish some um the results in this part and i want to
conclude that the light field sampling behavior is reflected in this model and all since
the spc preserves the vocal properties of the system is capable of on explaining high-level
behavior of the system like focal uh like um
depth of field or like a different phone rendering algorithms in different depths and it
is capable of extracting the high level camera parameters of interest and at the same
times if it keeps it simple but it is a high it has a high
school level and well there are some future works and they are actually ongoing works
related to this part and we are trying to investigate existing camera systems write ups
as one of the major points of the system thank you prior intention
well in this is taken consider them as a single optical element but there is
no limitation we can do the other weighting it depends on what you're expecting from
the model if the um for example you are going for precise result from the
model or if you're combining two systems and you want to keep precision as much
as possible while you're spending more on modeling the more complex systems this is a
trade-off and you will decide about how to work with this model but this is
explained basic behavior system and
oh
yeah and don't forget that we have and it's very sparse assumptions here you're working
only with um
geometrical optics and this is maybe event the worst i mean it's this is a
stronger something compared to what you are discussing
thank you