Let me let me know what I
asked you to be sure didn't I

love that book.

It's one of the great.

So this talk is

about the video summarization done by
my two students mentioned here and

just to show the problem.

Here's a short clip from
the tech elementary

or

I guess

on why we're here.

The good old days.

OK Few

haven't got the soundtrack
couple of thousand

I'm going to need teams working
in shifts going to take days.

This is basically how
analysis of surveillance

video is treated these days I
don't know what his assistant.

Meant but he said Not necessarily but

what we're trying to.

Dressin this talk is a little step
toward dressing this problem.

So in normal cases that
were a terrorist attack.

These are two example of
cases seen in Europe.

It talks for weeks at least
a couple of weeks until the video

that was collected from the site was
analyzed people went on the video in

they find pictures of the terrorists
here and there on the video.

The same and more recent thing
is is this case it was in it

was this was in Dubai two thousand and
ten and

again you take a couple of weeks and
people there underlies

the video found
the terrorist in the video.

But again all this time was
very long when he takes weeks

to find what's going on the people or
involved are long gone.

Something it's really different
wires in in Boston in Boston.

It looks to the F.B.I. couple of days
to find the terrorist on the video.

What was it so short.

So one.
If you look at some of the local

newspapers and others as rumors
Israeli technology may have helped and

then defy the Boston bombers.

Well that is very technology mentioned
here maybe the one I'm going to talk about

now.

But we have no idea if it was really
used in Boston is a client has this.

So use it but the F.B.I.

does not disclose what technologies
it uses in its Vista geisha and

so therefore I didn't acknowledge what
whether they did or did not use it in

Boston but the fact is the Boston
case was the first time that videos.

Were I never in a couple of

days rather than weeks to find
what happened in the scene.

Actually this is recent deployed in the
Statue of Liberty Empire State Building

has a places in China.

So there are some notable places
that use this technology.

Now let's go back to the problem
the problem is that we have millions of

cameras all capturing data twenty
four hours a day every day there's

more cameras installed all the time
each camera has more resume and

there's not enough people
to watch all of that.

And given that the person
can watch such video for

about twenty minutes before losing
your attention of falling asleep.

It makes their notices
even more more difficult.

So the result is that any recorded
video is effectively lost unless

there is some major crime or major
terrorist case were actually they take

says dozens of detectives work on shifts
to find the what happened on the video.

So the solution that we try
to do in computer vision to

this task is try to do it automatically
try to detect and track the object.

This is a big fan suppression try
to classify the object people cars.

Maybe recognize individual people
individually cars by the license plate

record now is the activity
that the objects are doing for

example a cliff like a street
fight the trouble of in this

case is there's a lot of progress done but
still a lot of more program remains.

Now even if even of all these
automatic analysis will

be very successful will
still would need manual.

Inspection the new inspection mean that
the people will look at the video.

So again the first two reasons why we need
people we need people because automatic

analysis is not perfect but the more
important thing is that the picture is

worth a thousand world and
the video is worth a million words so

and this will not change even if
automatic analysis were perfect

activities cannot be
described very accurately.

For example even there's not enough
words in English to describe colors.

There's order of magnitude more colors
that we can distinguish between than words

to describe colors not to say
shapes motion activities etc So

language whatever way the computer
analysis will be able to analyze

we cannot describe all the variation
of the nuances of what we'll see.

So people will always need
to see what what happens.

But again as we mentioned the too
much to see in man and equation.

Even after we've filtered we only
want to see police cars that

went in this road riding
more than forty mph.

There may be hundreds of such
cases in a day or a week and

then no one has the time
to watch all of the cases.

So the simple approach to do this one
approach is key frames based We have

a very long video and out of the video
we take a subset of frames this

subset of frames can be like
single frame every so often or

groups of frames or
smaller video clips to show motion but

it is based on on frames a frame there is
the main entity that cannot be broken.

This is a method that we are part
of it's a measure of object.

Based we don't care as we care about
objects we do to contract object and

then we show the object and once we show
objects we can play with time as you would

see in the in the taco especially we would

like to play object single tenuously even
if the object occurred a different time.

So in the basic video synopsis
we take all the stages of

the analysis video analysis and we assume
that only the first stage is done this is

the object detection of tracking and this
is not a difficult problem when the cam my

stationary when the camera is stationary
as most surveillance cameras are anything

that moves is an interesting object so
to find something moving in a special

camera is the easiest task that they
can be and we assume that this is done

and we take this trip isn't
it in the brief time so

that people who will be watching this
object will do the understanding

this is so this can show the results.

So the result of your synopsis is
the following on the left is the original

video is a video that
basically one day long.

You want to see the entire
video of course and

on the right is the synopsis
is a summarized in a minute.

All the object that happened
during the day are shown to gether

in a brief time you see that
it's very very short but

very crowded on the object
a shot together so that when I

give a talk in Bungalow people say but
this is how bungalow airport looks anyway.

Actually when I took off from
Bangalore I looked out of the window and

they were right.

This is how bungalow airport look
anyway it's a video on the right.

So wall.

What we present is a fast way to
browse the video and also index.

Because once an object is of interest.

We can actually point to this object
every object has a time in the original

video can go to the origin time and
show it and

the idea again is that object at the piano
region a different time would be

shown symbol tenuously How can this be
used it can be used in in several ways.

One is I have a budget of three
minutes before going to lunch.

I want to see everything that
can be shown in three minutes

of the most important object you can
turn three minutes another thing.

Show me all the object in as short
period as you as as as possible.

I can postpone my lunch of couple of
minutes if necessary and basically

this is only a presentation layer and
before it starts to talk about this.

I'll talk about previous work
we did on time because showing

object from different time
similar tennis Lee is is

in a way screwing up time so
time is not chronological anymore and

the idea of of basically not keeping

the chronological property of time
happened in an earlier work that we didn't

was aching to go and look at most liking.

If this is a video and
we'd like to Mosaic this video and

this is a normal mosaic
that will make up a video.

What's missing in this music.

It's the wind the wind is
missing the wind is missing.

It's blowing Can you see that when blowing
in the branches and the leaves and

the first we do we don't see it in
the second video and it would be nice

if we could do this video on the right
where we see both a mosaic and

the wind and

the people who are doing music and
for many years and

no one did this music on the right and
and there is.

This is simple in order to do the mosaic
on the right something wrong happens.

What what's wrong.

If we look at the shadow on the ground.

We see the shadow underground
several seconds before we see

the branches that cross this shadow on so

we see now in the shadows several seconds
later we see the branches on the right.

We see the branches and
the sugar moving singleton you speak so

we know ordered to do
this we must screw time

then cannot be kept cannot
be kept could logical

anymore and and this is basically

a kind of mental block that we had
to overcome in order to do this.

This was aching and once you block it.

It is easy to take it to other places.

So how is it done so assume this is
an original video this is a video

with a camera painting from left to right.

These are it was a falls between Brazil
and Argentina are said to be the largest

around that actually a very tiny
portion of these the really huge

falls and and
if we take this video on the top.

We just do it in the space time if
this is X. and this is the time

to see this one frame another frame so all
the frame and moving from left to right.

So this is where you would have this field
and the most like you saw on the bottom is

basically taking the right parts of each
of the frames and stitch you get together.

No we say well let's do another
music not take the right part but

still slightly lift to the right then
take the center of it take the lift part.

So basically have a set of
mosaics made of the same video.

Each time picking at different parts
of basically if this blue line

represent this part in
the original frame that we took.

For the mosaic we have a sequence of six
each We're taking a different part until

the last must take the left but now what's
the different between taking the right

part and the left but
the right part is just women a piece

enters the mosaic and the right part is
one the piece goes out of the mosaic.

So this is the first mosaic will show
an object when it entered the mosaic and

then it grew until the object goes out of
the music and and if we play the music.

We'll get this one this this video.

So we see in this video we get the kind
of Rama we get this water flowing and

this on the right a flowing and

we see water flowing together even
though originally they never happened.

Together and the reason is that if
this is an image we saw in this image

object in this time and the earlier time
and the beginning of time we see all

the different times we show them together
so this is how we can manipulate

this image by showing together
activity that were in different times.

So let's

look again at the video
that we wanted to move to

change the camera from painting from left
to right to painting from right to left.

What will be the natural way which
is the camera printing from right to

left leg backwards.

So here we play it backwards
the camera pans the right to left but

gravity works in the reverse.

But where the with the right thing
here is bending the camera for

right to left and the water flows down so
you see that we can.

This is a non-trivial manipulation
of time where the camera

motion has the right time.

But the drops of the right time by
becoming motion has a reversed time.

OK so how do we generate into the mosaic
we have a set of set of frames we run

many fold through the frames and
every pixel that the manifold touch

are collected into one image
when we want to see a movie.

We move the manifold and
this set up now generates one frame.

Sorry.

And this set up another for

us are basically a set of one frame
another frame and yet another frame is

in every frame is a collection of pictures
from this picture from this spacetime.

So if for example we take
this manifold this money for

these basically one frame picks
one frame and we chose a many for

them each times the manifold his
collection from a larger number of frames

and and each one of the many forces
a different a different picture.

For example is money for just collection
of these pixels that were touched by this

frame so when I say a manifold
touches a pixel it's it's

an approximation because the manifold
goes in between pixels so there's

a lot of thing to pollution going around
here so it's not actually touching but

I will ignore the entire pollution
that goes on here but it's not that

difficult to realize what the pollution
is so one example is good then if you

see the demolition of the tower in front
foot couple of days ago there were there

was a total of thirty two floors in front
for that was demolished two days ago.

Well the reason I mention because there
isn't one company probably the same

company that demolish is huge structures
in the universe and they have a website on

which to take pictures of their creations
so if you like to see destruction.

This is the place to go.

It's really thinks of the slow so
this it is taken from the website.

There is the King Dome in.

Seattle.

So it's here is the monish it's all in
the uniform was the fill down uniformly

if you wanted the don't to fall first in
the walls later or maybe they were first

of the don't later we cannot take it
again because the truck sure is gone.

It's just but we can do this with our
stuff we can take the original video

this is looking on top at
the spacetime volume this is the X.

direction this is why and will take
this many for what does this matter for

the Center at the center would be from
the future compared to the side so

the side will be in an earlier time in
the center would be in the future time.

So basically everything that we will
see in the center be in the future.

So when we run it through the spacetime
we see that the dome will for

first and the science will come out
later or we can do the opposite.

We can do the center from the past and the
side from the future when you run this.

We'll see that the sites will fall
first and the dome will for later.

So this second manipulation
can be done with this but

the manifold doesn't have to be
static it can be dynamic and

let's look how we can play when kind of
first sports event if you look at this

women competition but it's not clear.

We know all of them are the same but
we want to declare someone is a winner.

So we take this guy and
we have his style running faster so

if origin we start with
the original frame.

He stammered go faster.

So as we go through the time he

gradually would take for
the future in his winning the competition.

Well if this is a photographer
photographer in high school.

He's going to serve the video only to
the mother of the winner right no other

will bite.

So he want to sell more than one video
then he said well it's like another guy is

a winner and this one is from consented
to his mother and he goes and

creates another winner
from the same video.

Actually wanted when we went with this
Panasonic one of the recent Olympics will

give them a recording video recording and
video for talk a few of the Olympics we

try to offer them so that the broadcast
to every country that they're at is

the winner but they refuse a No one
will believe with you after that.

So they didn't get that.

OK now we go back to
the surveillance tons of cameras

camera did not look as ugly now
they look much less obvious

much smaller but then I would can
show them how many cameras are here.

No cameras.

There's one camera that being recorded.

But.

It's in here but
they're not fixed in the world.

So what I'll describe
here is me basically to

come under the screwed to a wall in
the movie and this is the story about this

work is that we try to summarize home
videos and eventually we declared

failure because we couldn't do home with
his we could do on the station cameras.

So we're telling around our failures this
sort of what do you mean it's a failure.

Most of video on Earth is
from station the cameras and

said I don't know I never take
video from session so you're not.

But look at all the videos around
you taking your picture and

his was right there much more video
taking from cameras like this anyone

guess how much of the video on
earth is taking by such cameras

something like think ninety five percent
of all videos taken by camera like this

stationary Actually it's even about eighty
percent of the disk space in the universe.

Not only there are a lot of videos but
the take all that this space because

think about such a video is how many
bits are generated every second.

Maybe has cameras so we once looked
there was a company that makes storage

so E.M.C. said that most of the storage
on their devices are bits of.

Surveillance video.

OK So more and more cameras used and again
not enough people to watch the video.

People say the professional
guards are watching one percent.

I don't believe them.

I don't believe that even one tenth
of one percent of being watched and

again everything not
no one is watching it.

So what being done so filled with this
video summary and every frame every walk

the key frames or a collection of short
sequences or adopt a fast forward.

Sam it sound of thought to do more to
capture the motion of the camera and

there's very interesting work but done at
Microsoft basically very similar to ours

of publishing the same conference but they
also allowed space to change the trouble

with allowing space to turn no one could
watch the video we need some stability so

if a squirt with time at least we should
screw up space so we keep space in text so

what we do we have object base and
we manipulate the time of object.

We don't do fast forward of objects so
we can do for us.

Whatever we can do but
this is not part of the method and

because we change the time of object
causality causality is not kept.

And we should remember when we see
summarization objects don't appear

in the right order that
they appeared in life.

So there's no causality it's a loss.

That's what we lose when
we build this summer.

Zation and again and
shown you this before.

How do we do this.

So basically we have the space and volume
of a very long video we have object that

to Piri different times but in this case
the pin different tub a different place

a person is walking on the ground
the bird is flying in the sky.

Why not show summarization where they're
both shown similar tennis Lee and

since they each occupy
different part of the picture.

They will not overlap so
we just when we show the synopses will.

So both the bird and the perp.

Together and we'll show shorter video and

they had these to take to take video
surveillance the take and take the people

put all the object in a database and
then when we need to create a summary.

We just collect all the object and
show them similar tenuously

regardless of the regional order so
this is for

example nine hours in the original video
now we chose in thirty seconds so you know

I wonder we're trying
to minimize collision.

We're not going to unless
basically you can always say

there is a in the minimisation has many
element one elements of what is the total

length of the summary another reason.

How many collision you have and
you can play with the weight and

collision the more weight
you put on collision.

The more sparse the video will be.

Yes collision the frame no knowledge
Emitter Theory reasoning at all.

OK So what are the steps one is
the take to interact object so

in database then select the relevant
object of the for database selection can

be by time show me anything from nine
o'clock to twelve o'clock show me all

the red people etc and then we
display them in a short video synopsis

where objects from different time
appear similar tenuously and

then make the data like time color
motion can be used to sort we can say

should sort me the object
from fast to slow or from and

we can play a lot and
this smells like a Google search.

Basically you want some features I want
read people's social the rich people

all come first and then the less read you
have you come later in the search and

when you find the person you
want to stop and of course.

That's always a way to index
when something is interesting.

There is the time or age of time and
we can play play they

play the original video so
that the text of tricking is done by

something called because subtraction
which is very very common

every pixel basically is examined over
time and for every pixel is statistics of

the color of the picture is being kept
like a histogram mixture of God etc and

once we have a statistical model of
a pixel The minute the pixel a right

not a pixel a color after this pixel
that has a very low probability given

the the probably stick model say
OK this might be a forgone model.

So we say at this point.

This will be a foregone model so
every every pixel has

the probability to be in the former
model and then some kind of cleaning.

We want foreground object to be in groups
because object a more than one pixel and

we want them to be also
consistent in time.

Don't flicker So
after some cleaning and collection.

We get the moving all
object in a frame and

also we do tracking in time there
is actually a nice library for

backup subtraction methods in and

a kind of Google Code and
some of them give nice results

then we also use trekking
we detected the object for

going object in one frame another frame
the next frame we can track them for

example if we want to see a group of
pixels in different them to be an object.

They'd rather be similar to each other and

this is especially important when Object
cross each other so want to make sure that

the object that enter this occlusion
we continue with the same object when.

Switch to a different object so
we use shapes similar for

tracking we use common field to make
sure the motion is similar nothing

special here and we end up basically
having we call a cube stupes our

object a long time and this is how what
you look like on the top is a video.

So the illuminated part is the part that
is part of an object in each frame and

if you compress the time
into one object so

this is all the instances or the different
type of an object as shown here.

So basically an abstract object is
something that you'll be in in spacetime.

Then the picking it we have the set
of cube in spacetime these are again

from view from above
the tube in spacetime.

We want to pick them for
two purposes one purposes is we

want to pick them to get shorter video and

we want to minimize the overlap among

the new we don't want collision
among object in the pic video.

So it will get basically there is
a cost function the course function and

basing the temperament because time
shifted to vote for every object.

What time will be the time it will
played in the in the resulting video and

this is kind of to show as many videos
in as small a time and then we have

as much activity as shorter time and here
we have coalition cause we want the object

not to collide with each other but it
doesn't have to be strict because sometime

an object can be a long time and
we would like a local Asian and

this is temporal consistency for
the leaf we import for

us that object that appear earlier would
appear earlier of also in the summary.

Then we can put the larger.

Weights and find better the people
we can play with to get

to get the different summarization and
actually the minimisation of this

function originally we did it
using simulated I mean it's

a very difficult problem but when this was
turned to a company the company said well

simulate an inning doesn't work it's too
long people wait to look for the summary.

But it's some kind of
greedy algorithm that work.

Almost as well as you will and you know

what the videos were a little longer.

So I see where there was some some kind of
coalition want to tolerate if we do see

where they're leaning we can get a shorter
video with the same amount of collision

compared to exactly that.

But of course you know.

OK so let's give an example for

example this is a video this is what still
didn't do the reverse of the Pause for

video summarization So there was
a student coming from the left and

then some come to the right and this girl
this is the she's coming for a long time

coming and going and this is a poor
rigidly Someone come from the left and

disappears someone come for the right and
disappear and then for a long period.

She's came in and
left basically to do this

one to do it to shift this
person in with then this

put this sideways and the summary is
shown here is that she is coming in.

Well she's taking her water
everything else happens and so

basically show everything
in the wrong order.

But in the much shorter time.

So what the show now is we
have object at the pier

ocurred different type of
piercing maintaining Asli and

we can take object that appear very long
time and break them up into pieces and

this is an example if anyone knows
Michael Cohen from Microsoft.

That is Dr Linda and he can look at this
for a long time for us it's too long so

we want to try so here she is in the one

third of the time we see all she did and
the monkey on the monkey bars.

Again we can look camera special

card when a school to the one of the color
of tape we can consider the stationary.

So this for example in the mosaic.

So we have how many copies are here.

It's funny it's hard for me to count when
it's moving when the stationary sees it.

But when it's moving it's four or five.

OK then we have to create a big round we
have to place the object on the background

or region if this is a short time if
it's five minutes into five seconds then

you can still one still image can be
a big one but whatever to do a full day.

If you do full day so we have basically
our background is some kind of a fast

forward of a real big run for
example this to be a back ON THE because

it is all the object
without the moving object.

Now how come we see color that we see
car there because the sky were parked in

the morning and
then left in the evening so

we do statistics of the of
the big round after

some time few minutes parking
the car becomes part of becoming so

the background does change does
change with the amount of parked cars

that change with the illumination and
this is a movie showing.

OK So

this is a bit on showing
the background just the background and

we see that we run faster.

Night because if you are frames at night.

So this is the background video
on which we place the object.

So when we Person object we find out the
first thing that when we put object for

one time on a bit one from another time
we see a boundary like you see this

box around the object because the object
was when the ground when the background of

the brain hit the brick is brownish and
it really kind of disturbed the viewing so

we'd like to blend we found that the
person editing kind of will do the best so

basically takes this blending
go if we take this background.

This is after porcelain if you see it.

Each object here comes with its different
background another with puts a look.

Get the right.

So it's stitch well.

And if this is this is

a big this is kind of one of the videos
this is the the the synopsis.

If we can let's just say
look at the synopsis.

This is one frame of the synopsis.

We have three object we know
that the three object for

different time we know this by the shadow.

So one object is a long shadow probably

from the morning on
the evening another one.

No shadow at all.

Probably was cloudy and another one of
those should wonder if probably at noon so

but we know that these three objects
never saw each other in the real life and

but when we want to know what happened.

We can select an object and and

it will point to the original

video where this object was shown kind of

to use cases that I like to
show is one use cases is super.

A market and basically a car here.

What was hit and the client

came to the security officer look you have
all these cameras in the parking lot.

I want you to find who hit my car and
he asked How long

were you in the shop set for an hour and a
half so I'm going to watch for an hour and

a half of video then you realize that just
a week earlier they bought this video

summarization said OK you went to
the video summarization summarized

this hour and
a half into about a couple of minutes.

And here is a playing of
the summarization look at this car.

So you see you know boom
someone see the car and

then goes and park somewhere in so

it was very happy that after twenty
seconds we found rather hour and

a half we found the car the hit that ran
out of the parking lot the car that it was

still parked there waited there for
the driver and exchanged the details so

it's interesting that it doesn't really
have to do terrorist doesn't have to do

two weeks in can be one hour in
a parking lot of a supermarket.

Then if you want to spying on your
employees when they go to lunch break.

You can take a very short video and

see who goes where you see the collision
effect this guy with the green he stand

there for it all the time.

If you want a local lesion
then we could summarize it on.

So we're localism that no way so
you know they

don't talk to each other is
waiting there for someone so

this is an example of a summary and
therefore now we watch for

your collision OK we'll just skip the.

Actually one of the thing is the measure
similarity is a similarity can be in shape

and trajectories of it's obvious.

So I like to skip.

OK So what do we have a lot of
major data that we accumulate on

an object one make the data is obvious if
the color is the trajectory is a speed.

It's set of safety features
that we have on an object and

another is less obvious it's something
that we need to do is if the person is of

the car if it's a person maybe
what's the name is it the man and

man or woman is it a hundred or
is it a Toyota.

So everything goes into
what's called meta data

make some kind of symbolic information
that goes to the object and

then we can then select based on this make
the data what we would like to see and

basically the best thing is to run
the closest store description first and

then less so
there's an example of a video made at some

percentage of any university I don't
remember which university this is and

this is the video mostly students are
crossing from one site to another site and

again if you watch it you can
fall asleep very quickly and

assume you want to do summary.

This is the synopsis so this is a summary
we see all the object together and we

look carefully you can see the problem of
object the texture of tricking you can see

objects kind of disappear in the middle or
not but again it's very organized.

When you see we have something in mind for

example assume we want we want to
find the fastest object first.

OK so we get the full.

This.
Nope sorry.

So what is the fastest first we see
the bicycle are the fastest objects and

the little what's the next
fastest after bicycles

and then to go father sends bicycles and
the next.

It's running so we can sort by
speed until we find out what we

can sort by color for
example to the Reds or

first we see the object that there
are a lot of reds and then we see less and

less red you'll see bugs for example
you'll see some some with no read at all.

What's doing there are not true.

See here these two girls
have no rent at all.

So good that whatever Google
sorting is worth something.

You see you don't know what coming
from we can also see a similar to so

you select an object and see pick me
show me obvious similar to this you know

guess what similar similar
is the trajectory.

It's a is that the ration
the direction and the speed and

also a set of sifts that are there and

the first object here is the object
that we selected the similar and

then we picked up all the girls with
shorts and go in the same direction.

So now maybe skin color was
the dominant Of course it's

natural that the speed in the reaction be.

The same but also maybe the skin color and

it's a few Pruitt longer we'll see
man we'll see a little other people.

OK so I'll end here and thank you.

So the question is if we can analyze
the money political video and we

accelerated once made to the other and
find if this was done

and most of the time I think the answer
is no if especially the swimmer what

happen is that we select how do you do it.

You do it.
Basically you do it with two things where

you do you do graph gets to cut in
the places that you want to feel and

then after you do the graph got you do
some kind of blending so there's no way

you can do this now if the cut his
to go through objects for example if

the cuts when we do this we met we did
the first the cut without graft cuts.

So we were cutting the middle
of a hand of a person.

It means that you do like this.

So you take this part of the head
from one time and the other part for

another time then the hand will have
a funny shape because it's the NOMIC But

when you do it with graft cuts very
tough very hard to find anything

actually one of the one of the best

reviews a God of the people came for
us to do for Encik analysis of videos and

says Can you guess videos of krill your
videos are most difficult to analyze for

is and for us if you want to see if
we have manipulated the video and

created at the fish we did.

That's not correct then there was a new
reader the most difficult to analyze

forgive provide them with samples so
that as a compliment

of what went wrong.

I'm like yes

yes yes cellos is a.

Big problem.

And of course the big thing is to say
well let's ignore shadows right now we

don't ignore shadows because we don't

we haven't done it with
a warm motion detection on

object attacks was only made on
the fact that the object moving.

Of course the most important on is person.

So we'll try to do person recognition in
which case we could ignore the shadow

right now we don't know the shadow and
you saw in the in the example of

social we see three objects at the same
time with different shadows and

then you know that the three
object in four different time

even if we show the object without the
shadow you know object have said to write.

How can we have object without them.

So

it's

we just tell them that they
should remember that it's

time is not correct and
that So now I must tell you

that I never look at these videos right.
So

I don't know what the reaction to
give to people who watch a video and

actually one of the things that we
didn't realize is it helpful or

not maybe by putting many object
together just confuse the analyst and

feel better to see one object after
another rather all of them together

we don't know the answer but
what the F.B.I. did the F.B.I.

was very interested from the beginning
in this so they took the softer

the plate with a couple of months.
They didn't take what the result is but

they bought so perfect that they bought
after a month and means that they found

that it may be faster to analyze
the video after the summarization

than the before but they don't tell me
then they only get information they never.

Volunteer to say anything even if
you ask them their views to say but

basically when there is no power locks the
issues paradoxical when you have products.

You cannot tell between a moving
object and the parallax motion.

Now if you have object recognition.

You say OK I want people so human you will
find the people even if there is part of

the product or not confuse you.

But then the question is on which copy and
the pilots will change the background

check which I show in which copy of
the back and we'll put it so it's not that

obvious so we're trying now are working on
wearable video I mean where the video and

I NOT mean the student where videos and
that they walk around and

then the question is What do you
do when we can summarise this

using this method when you walk around.

Maybe when you sit in the office that and
students are coming in and out in and

out you can summarize this but

that any other more interesting
activity is more difficult to summarise

actually if you can do three D.

positioning then you can plant
the object in the right thing and

the really good my building room in the
day you can put the object on a three D.

model but this is much too difficult.

I believe then for

something can be done quickly used
a summarization we need to be done fast.