thank you very march for introduction um
what didn't have brought a i'm very glad to be here and have possibility to
present a little work which would on a nice place and interest if a cell
in germany i continue to speak about uh a human head reconstruction and my part
of these beak B project uh which was introduced by my colleague uh is stitching
of reconstructed so face might focus bouncing the variational registration of for each range images
uh with no by non rigid deformations so as i said this is a part
of the project where we want to reconstruct human head the model if you look
at a recent set of images
well my colleague expand already uh but
what happens after the reconstruction all parts of human heads reconstructed separately and they are
uh they should be matched and then stitched together in order to have
and model which is you use appropriate for printing a for us it means that
we want to estimate and model which are smooth and that and have no but
see johnson and staring effect in the place of station also faces
the first part of all for all of work flow is done by uh structural
optical flow algorithm is already explained this uh the result of this algorithm looks look
yeah results billboard but unfortunately they are not perfect a they are subject to distortions
this distortion scores by
many factors uh kind of the economy's calibration and some errors during reconstruction and all
this for example we use uh so called they have E much more range image
for representation of four so faces uh such structures can not uh described um occlusions
or uh or the discontinue two regions on so face in a proper way that's
why oh uh there are a lot of place also faces were some false box
i hear uh another problem another problem of the soviet it's not nonrigidity of object
of interest uh when uh we say to the model yeah to human been police
station don't move it doesn't work queue moves and one we make several uh several
uh shootings i would say uh we uh we have such problems that
several parts of this so faces for something have plotted more departs they have different
position one three or one for each of the relatively to the to the to
all the uh and it cost
it makes the problem of stitching very difficult uh because of such the did distortions
we cannot simply use existing matching the utterance uh to solve this problem uh icp
for example integrate iterative-closest-point generate a result which still uh has a lot of places
where so faces simply cannot together so for example here it is green part it
is uh
to occlusion uh this distortion is generated by occlusion here we can see false matches
because of discontinued see uh see it just problem of for a couple of the
small errors and kind of calibration and so on
so we need some additional deformation local non-rigid deformation which can uh
can improve this case can compensate these distortions
uh the main the main challenge of my work is to combine global rig it
uh transformation of so faces with local non-rigid transformation also faces in order to bring
them together and speech
so i want to explain our methods just promise simple artificial example here so physicists
represented in two-dimensional to two-dimensional manner uh so all i had to find a so
face which will be similar to both so faces which are not which are not
so which in intersect partially which costs move
uh
so what to do we search for a some transformation or some transformation to which
transforms of faces in such a way that we protect smoothly full to show some
people to show that means that you in your of these so faces is a
small suffix
um
so because of
unknown comfortable in
come complicated distortion we can not consider once a faces a template that's why we
uh we formulate the project of make a proper problems matching not it's problem of
matching once for face to another but as a problem of matching what's the faces
are also faces uh to some expected so fixed what is expected to face expected
surface is our target sort of experts so
we say that we want to manage our suffices to some expect so face which
is in principle our current guess and box resulting so face about are just suffix
all this so expected to face can be
can be estimated by using a simple consistency constraint which says that are all points
of also faces should be consistent with
consistent would expect suffix
the mean means that minimisation all day so constraint bring cost so
to the
to the equation for its estimation it is not all the than simple uh weighting
weighted uh some of are also of our source so faces
he a weight is a quality measure for each the face which got comes from
reconstruction algorithm i will come back uh to this quality measure a little bit later
so ah S A I C S I
already mentioned we search for some transformation we which is a combination of some rigid
global transformation and some non rigid local transformation more global transformation should bring so face
part uh in the proper position in some common uh coordinate system a local transformation
should compensate distortions in order to bring them to get the question now what kind
of local transformation to use we can see the so called sinful well as a
prominent way to compensate such deformations ten flow is a is a three dimensional of
vector field which describe this uh movement of corresponding point in space it is four
dimensional and a lot of optical flow and can be can be estimated in a
similar manner but we consider a variational approach for estimation as
a signal to promote its among
among colors
so
that's why we formulate our matching
so that our matching uh problem as in the term of variation requires of appropriate
energy function which should be minimised and this energy function consist of set or combination
of data constraint constraints and combination of regularisation constraints
we use
we use free constraints it they are toast data constraint which says that a lot
transform it's or face
should be consistent with expected surface
a smoothness constraint is used smoothness of scene flow constraint is uh it is very
commonly used constraint for minimisation problem uh it says that the gradient of target function
should be mean you should be more in order to simplify calculation will know is
also cool common approach we approximate this gradient aspects the difference between what will you
the function and its
mean some mean or some approximation of the function in some neighborhood
at around current or
and we also use so this so called tikhonov regularisation constraint which says that uh
our scene flow should be as small as possible uh
we use this in order to with situation that scene flow describe some uh transformation
which can be described by a global or transformation
so
yeah O combination of these constraints
give us a common energy function for both transformation for jane transformation here we also
use additional weights for each constraint which uh
which required for
oh
for making it to make a reconstruction process more stable to noise and more uh
more
contour again
i will explain it will be this uh weights
later so the solving of these of these um realisation problem
because i'll energy function is a function of
many parameters of two to twelve global transformation parameters of for local transformation we search
all this parameter separately first of all the fix scene flow in some yes
uh insight in some point and star which is our current guess about seven flow
and
received energy function which is the energy function of only uh global motion parameters this
formulation after this could decision is a
a weighted icp uh i would say uh weighted icp and it can be straightforwardly
still so that the
the result of minimisation of this energy function is used them uh to
to receive energy function for stencil so we iteratively uh optimize
well global transformation parameters and then local transformation or
um the
this is the uh this problem can be solved with using a or do not
let me articulation uh will bring cost to the iterative solution iterative algorithm
after receiving of result of both these all of minimisation able both this function we
refine our expected sort faces and should start this uh procedure again iteratively
so some words about
weights the weights needed to
to reduce to regulate influence of if of each constraint in you know energy function
uh depending on cable on quality of
uh source data which are involved in this computation um each source or face to
after reconstruction are is supplied with a quality measure in you know
uniform of weights which two
depend on noise level of
of not of so face estimation yeah
uh the main idea the main idea just
if we have a good estimated uh so faces
it means that
both correspondences
are estimated it with high precision
in this case weight of data constraint in
energy function should be higher so that the gain of data concern should be fine
otherwise smoothness constraint should to should get a high higher you in and so let's
a smoothed or bad data
um we use to go common uncertainty propagation uh rule the to estimate the weight
of the data constraint but see that the noise level of data constraint is the
sum of noise levels of that are involved in this data constraint just a lot
so face and correspondent particle for expected surface regardless of
a you can have regularisation constraint by weight of tikhonov regularisation constraint we can uh
kick can control the growing all for scene flow from the T control the uh
the power of local transformation um conceptually that we prevent local transformation in places where
so faces
are estimated with high precision is if we know that's a face estimateable high-precision would
don't want to bend it
that's why we use uh weight of uh so face which should be transformed as
uh wait for tikhonov regularisation constraint
and uh wait all for smoothness regularisation constraint not now we calculate this as the
sum of
as the sum of form weights um of sinful or which are you stand for
weight values which are used for estimation of mean value for regularization for a for
a approximation of gradient and to be honest this question
question how to calculate
okay it is still open question i just want to show just some results so
here it is a distance map maps between two so faces which we uh mitch
with icp algorithm and propose it expected to face mentioned
uh so this is initial match which was done manually the result of
icp it looks like icp uh tries to optimize global distance between so faces
uh
although it will have um has
tendency to move for so faces
one to another in places there the where they are more bare feet to each
other that we have more consistent in
for example this part of face and parts where uh we have
problems words of basis the form of this
uh these parts are uh if you look more closely it happens because in this
place is expected sort face
or let six four paces some more can a consistent and expensive face uh receive
more weights
so and application of scene flow remove practically all these deformations
oh on expert O one which is a fixed i just i just want to
show you three-dimensional result of this much is not perfect because there is still a
lot of problems first of all we cannot uh automatically this uh find difference between
uh between false so phase parts and so on so face parts which have no
uh correspondences at all box in places of where so faces i just the deformed
not that much i don't want to pay a lot of attention on these artifacts
because it is still subject to of work but in public in places versa faces
ah
not very strong deformed we can bring them together and stitched so for comparison i
just show how it was
so this is the initial match
here so faces you can see that these parts
can not be measured by some documents so i switch to
two
oh
what's
okay thank you very much for your attention and
well
uh_huh
oh
to be able to be honest yep for sure the lighting conditions it's very oh
how to see printed they inference very much to reconstruction but uh i had to
be to be honest it is it is not question to me because this reconstruction
and it was presented before uh
i
what do mean what it is possible to find the position of the light
more than just this problem was not can see that but for sure it is
i would say it is another fields of fu
which one
i mean whether we use a touch information for this matching
okay
no we do not know such information with the feature description uh for matching only
uh that that's information is involved in you know matching strategy on the in a
way that uh we add additional weighting factor which is that a way we will
correspond