uh i think it is my pleasure to present my work here and then a
nice young channel and um now working in france at C inside your uh the
subject of my paper is kernel similarity is that active appearance models for face recognition
uh first of all i'd beginning uh i want to use a few words um
active appearance models which is the base of my work uh it is quite to
the widely used to yin uh face recognition and the object tracking and sometimes for
uh medical image processing uh they in sexual uh the in central idea of this
matter is to uh build a model which contains pose the shape and texture information
uh of the training data uh and then when we have a new face to
recognise the model can generate a pen generates the
appearance of the new phase of id three shows so uh but us using this
matter that we can get the detail of the appearance of a new face and
also the uh location of the landmarks on the ball boulder of the base and
the dog runs uh so it is still quite powerful wow algorithm for sometimes for
the uh space
for the recognition of the face expressions except for uh and the this matter it
has it's lame eight so one of the most the important to a problem is
uh is that the quite sensitive to the illumination conditions uh that is so why
we want to our improvement
and in recent years there was quite a lot of researchers work on this problem
the illumination condition and the uh there is some idea is so what are some
of them the posted to at some action a some action no parametrise scene this
model which is the red delay corresponding to the illumination and to decide he rates
they depend it to this work by using this site here uh active appearance model
is able to generate the face is in a quite dark the illumination or a
very bright illumination but uh we know that the illumination that is not the worst
thing for the illumination when the uh illumination come from one side of the face
it to make that is half dark and half right and this is the more
complicated thing and the this idea that in the work for this case and someone's
supposed to uh apply a filter on the decoder for two can uh invariance condition
uh for example uh for here a transformation all couple filters uh and this matter
the also cost to lose some data use the information from the recognition images and
there is a kind of quite a tradition idea is to use some other october
no uh transformation instead of principal component and that is to extract the uh the
most important variables uh in that database and the although it is quite old about
so we believe a we find the and up operate the transformation uh it can
work for this case and this is the initial idea of our work
and in this page i want to presents the database aware working on it is
called a cmu pose illumination and expression database of human faces and descent also in
this data database is captured the in section and environment uh in the room there
is several cameras are wrong that the volunteer and several flashes are around him and
to make a different illumination conditions the flashes flash one by one to get the
illumination and the here i give out some examples in this database uh for each
person your uh us fourteen different pose of that is and for each post there
is twenty different illumination conditions and we can see from these pictures some of the
illumination it's quite complicated but complicated and heart to recognise
one so we decided to use this database uh with data statistic and the light
on the database uh here in this page the we show that histogram of the
euclidean distance between each of actors uh of the uh each vectors O and so
one is for here is the uh that terms of the shape and here that
is the distribution of the texture vectors uh we can see that so for the
shape vectors it is the close to a gaussian distribution and for that extra ones
uh it is so quite beautiful passion distribution uh this result is quite interesting and
the according to this result we decided to use the kernel to be able to
a similarity
uh similarity matrix the instead of the covariance matrix which is to use the in
pca uh have to see that it's uh occur no it's not such a new
id reading in this case requires the kernel pca came out to maybe twenty years
ago and it is directly used to eating active shape models which is the a
priori is work of active appearance model but that the alternatives and to continue to
use it in active appearance model requires it is very complicated to uh reconstruct the
phase and we construct that you major from the uh extraction features
but uh for the active appearance model it is very important to reconstruct the images
um here the proposed method we call it to a kernel similarity component analyse it
is quite different from the kernel pca but uh sometimes it seems quite similar with
each other and to mathematics star race is quite so it's clearly reading you might
paper he i don't want to uh talk about the can uh mathematical conclusions uh
just the procedure of this matter what uh is very simple which is to use
the kernel oh
we just use the kernel to build the uh similarity metrics and then calculates a
the principal component to from the uh
from the uh from up from the metrics it uh and then we get the
eigen faces which uh which represents a to the most important variation in the database
uh and here in this in this page what i want to this i want
to show how the eigen faces a facts to the variation of the model on
the left part is the uh result of the proposed the matter uh we can
see that for the first and the for the third feature uh it is obviously
control the illumination variance uh the illumination environments on the face but so for the
principal component to analyze um
the bar uh the variance is only between the genders are sometimes between the different
to uh shape of the face
so this result is tell us that so we have already choose the appropriate the
transformation um and here is the experimental results on the right C is the a
bit evaluation curves which what i don't like it's uh i like to see directly
the
the image is as we said before for uh yes the first column is the
result of the proposed method and the in the middle column eight is the result
of standard again and left column is the original you make use a which is
a to recognize and a as we said uh standard and it works well when
the illumination is not that complicated but so when it is how dark and half
right is uh the a and that and work but so the proposed method it
gives the quite good result
um we also applied this method in the a rotation of the phase in this
problem but the uh from the U matrix we can see that the improvement is
not that of years uh only for some certain case it's some uh some change
but not a lot
uh and here is the conclusion the proposed kernel similarity is the active appearance models
is robust to illumination and pose changes of the pc images with this signal them
at their depict the fitting procedure can accurately thing sizes bases for my right to
my dark affected by the illumination and say i have to emphasize that this method
has the quite big limit that sits requests uh applied set accuracy at a alignment
of the shape and texture vector if we couldn't do this it won't work would
and so that the next step for our work is that so we want to
make the matter to work on both the pose variation and illumination variations
and that this all thank you
estimate the parameter of the
so i can understand you clustering
yeah uh_huh
is your rights yeah there is a
i you mean you mean i think you mean this one yeah this one oh
yeah this is also part of our work
you see wow build a histogram of the uh of the mecc vectors and so
we uh and to that sir it shows the gaussian distribution so we just use
the uh
variance here
it is a portion and to the kernel we use is the portion so
we think here that it that is it's represents the variance of the gaussian distribution
so is that the
uh excuse me that's what is the other uh algorithm called you see
uh yes i think i heard that before
yeah actually the principal common to analyze is quite close to the uh independence the
common is except that that's a we try this matter what right do you depend
and one that's it doesn't work quite good
oh yeah
uh_huh
yeah actually what have uh convert opens up the uh ninety eight percent each of
the information