
[Music]
[Applause]
so what you see here is the luggage
built in an airport it’s usually the
place where after a long flight you
expect your luggage to arrive if you are
lucky right actually it turns out in
airport serving around five million
people per year around four to five of
these people decide to take a joyride
on the luggage belt this is actually a
nightmare for security they want to
detect it they want to prevent it they
want to find a way to be able to solve
this issue and we usually are the guys
who are involved to try to solve these
issues and similar like these I work in
the field as mentioned before of machine
vision and artificial intelligence so we
have a team of people being involved in
trying to solve a real-time issues which
might be relevant to different kind of
customers or anybody else
let me tell you how it all started so it
all started back in 2011 we were invited
by railroad company to solve a problem
which was actually solved already long
time before so what they had was a very
simple issue every year around million
wagons in a bit less number of trains
arrives to a country what happens then
as they stopped in the first station
five train inspectors using
walkie-talkies jump out of the building
next with train station and they walk
next to the wagons and they say hey this
is number one this is a regular number
here these are the defects these are
things we find about it this is how it
works it was not really very efficient
process it had a multitude of problems
as you can imagine first of all it is
never a good idea to put the person next
to a 60-ton moving wagon as you can
imagine so that’s a safety issue
second of all it’s prone to human error
so people can make mistakes and you know
when they do so
walkie talkies when somebody writes it
down sends it over fax machines which is
really long time one historical
technology back to the center somebody
reenter this this is not efficient at
all and of course another thing is there
this itself is not really a profession
where you can expect really high growth
but you can’t really become the best in
it or do it the fastest it doesn’t make
much sense as you can imagine so we
weren’t invited to solve this issue back
then and after half a year or a year so
to speak it turned out to be first
failed miserably it turned out we
couldn’t solve the issue straight away
with the technology which was employed
in many places around the world let me
give you some of the reasons first of
all you get yellow numbers and white and
these kind of things
second of all sometimes I get the
feeling people who actually write these
numbers have a special hatred for using
templates as you can imagine there next
thing the vagon to the right it’s
supposed to be clean as you can see it’s
quite a problem to see the number over
there even with your eyes even with the
naked eye that’s another issue which you
might have so we were fighting with this
for a long time then in one of the
meetings one of the researchers who was
not the software developer who is not an
engineer said hey you know but when you
look at the wagons for specific types of
wagons the numbers are in very specific
locations so why don’t we use the first
identification of the type of the wagon
and then we know where to search for the
number which makes our job much easier
so this was actually quite successful it
turned out this way we can do this we
can actually help the guys to solve the
problem so after this we started
actually bragging about ourselves you
know we had this major success that
coincided with the time when artificial
intelligence and machine vision actually
became something like a little bit of a
hype around the world so everybody was
saying this is really cool we should use
this technology everywhere so we were
wandering around the saying yes we did
it we are really good at it we can help
you solve any problems whatever you have
it didn’t take long after a short period
of time we actually were invited by
subtlety to solve a problem they had so
they had two roads one of the road was
leading to a nice beach which in summer
people went sunbathing doing everything
else second one was actually the one
which was more practical that was the
one which was going to the next village
so people were using their bicycles
riding through the shops to sit the town
centre doing all these things and that
was much more practical so they had they
had a question which one should we
develop we said yeah of course we can
solve it that should be very easy right
so the solution we came up with at the
time being was very very simple we said
okay what we can do we can actually see
if there is a motion so we’re gonna
place a camera on a pole we’re gonna see
if there is a motion there if there is a
motion there we will look at the picture
and see if there is a bicycle on the
picture right if there is a bicycle we
say plus one everybody’s happy we know
the number of cyclists sure then we run
into this by some accident somebody
actually drew a bicycle on the pavement
in front of the camera so as soon as
somebody walked by that you know motion
bicycle plus one this was rather easy to
solve right so we trained the system to
understand that this is not an actual
bicycle right so actual cyclist looks a
little bit different and there is a
context in all the other things but then
we run into another problem so we sold
this one and then on a hot summer day we
suddenly noticed there’s a number if
there is a rise in the number of
cyclists but on the wrong road on the
one leading to the next town not to the
beach why so we started investigating
this issue and it turned out there was a
small problem with this as well one of
the cyclists had bought an ice cream in
the local town store and he was going
back home so he stopped somewhere the
best place he figured out to stop was
right in the front of our camera we
didn’t like the guy too much at that
point of time as you can imagine so what
he did he started unpacking his ice
cream movement cyclists plus one then he
he raised it up to his mouth faded
movement cyclist plus-one and within 15
minutes he did it you can imagine we get
like 300 cyclists out of nowhere out of
the blue so you can imagine there is a
special type of as we understood that
the special type of love for the
cyclists versus ice cream and all other
things this was something actually one
of our customers told us how to solve in
one of the meetings he said hey but you
know why don’t you just follow the guy
in the video you have the video stream
right
you just follow him as he moves around
you just found him once this was the
solution we came up with the one we
actually employ which turned out to be
much more successful so it didn’t get
interrupted by somebody eating ice cream
you know pushing buttons in the phone
and all the other things but not all the
stories we had were the ones with happy
endings so let me tell you a little bit
more about the ones we did not so first
of all but start with the one I started
with him very beginning so people taking
Joyride on luggage belt right there were
two ways to solve this the way we saw it
first of all we tried the one where we
said okay let’s train the system to
understand what a human or luggage boat
looks like how do we do it we don’t have
enough data it’s only four people per
year in the luggage about right so we
did we got our guys on a luggage belt in
different kind of positions as you can
imagine and try to create some data to
train the system okay
well didn’t turn out to be that great we
didn’t have enough data still and there
was still some errors so he came up with
another idea actually the customer came
up with another idea he said hey why
don’t we use the thermal imaging where
we can see the temperature of the person
even intoxicated person has a
temperature of 36.6 degrees more or less
give or take it turned out the issue was
that it’s not only that within the
within the all the two days of the
summer we usually get here real summer
suitcases turn out to have the same
temperature because they stand out next
to the airplane you know in the Sun so
that was an actual failure
so that and being really not happy about
having too many false positives or false
negatives this was a problem which we
didn’t manage to solve so this is still
an open challenge which we have then
another one we talked about that one I
don’t think you can see what is here the
the thing on the right is actually one
of the last images one of our systems
took in the small town of Louisville not
so small but still where somebody
decided that it’s well there is a pole
there is a camera why don’t you just
pull it down right so this is no machine
vision this is no artificial
intelligence there was a very simple
solution you can see on the right
you put these sticky things on the pole
and hopefully nobody else climbs it
anymore so that’s fairly simple but
there was another thing related
artificial intelligence and machine
vision which we actually had to solve
and which we didn’t manage the soul so
the customer we had they said ok now you
count cyclists you count pedestrians
everybody’s really happy about it why
just why don’t you count the stray dogs
how many stray dogs are there how many
you know dogs running around without
people you know guy like this really
nice yeah it’s not here mm but still
it’s a bit it’s date time the question
which turned out to be a problem was
that it’s not that easy to understand to
identify even by a person what stray dog
is there’s no human insight it doesn’t
mean the human is not nearby right so
that’s another issue which we have so we
were not really able to identify the
clear definition of the problem what a
stray animal means so how long do you
follow it how long do you see if it’s a
stray animal not so this is still an
open challenge which we still have but
along this way as I said we get invited
to solve different kind of problems
using vision and intelligence but there
were several frustrations which we ended
up noticing along the way and several
learnings which we got along the way
first of all as you can see this has
been a bumpy road so some solutions are
there some are not first one we got was
actually quite interesting whenever you
look at the research when we look at the
data which is presented public
you usually get things like saying you
know this is a sample data we used these
are the pictures and then you just train
the system to say to classify them
that’s fairly simple usually the way it
works it turned out when you have to do
that position of the images yourself you
learn a lot more about what are the
issues with actual photography I’m no
photographer I mean my wife always said
that the easiest way to ruin a picture
is to give the camera to me right this
is actually a clear example of that as
you can see I was trying to catch a car
not that successful yeah but still I and
our team know way more than we would
like to about photography I’m still
crappy photographer don’t ask me to do
any photos but now we know what much
more about what the motion blu-rays and
other things which are related to this
how to build the lighting system and
everything else are these kind of things
but that’s one more there other thing
which is important was research or
frustration many times in our team we
got young guys coming back and saying
hey you know we tried to use this
research which is publicly available
which is the newest thing in the field
everybody should be really happy about
it
but it doesn’t work and then we started
investigating why it turns out that
quite a few problems in that one
first of all in many cases the research
has some small hidden things so it’s
publicly available but some of the data
is hidden second of all in many cases
it’s trained to be used under very
certain very specific circumstances so
that’s another issue which which you
might have right
the third thing which we learned certain
frustration we had was the frustration
of the problem owner
so the problem owner usually was
presented with an idea that you know
artificial intelligence machine vision
is here to solve all your issues right
now it’s already here it’s beating chess
masters in chess your masters and go
learn to play games out of different
kind of things but when you try to apply
this to a very very simple real-world
problems you either end up having
something which is very very expensive
and high-tech or which doesn’t work the
way you intended to work so the
conclusions we had the things we came up
with in the end of the day
actually we’re fairly simple first of
all we understood that it
we need to deconstruct the problem we
need to split it up in pieces and this
is not necessarily where software
engineers and engineers in general
prevail sometimes it has to be people
who actually might have a different
point of view and things people who are
involved with these problems on daily
basis through deconstruct it into small
pieces and then you figure out which are
the areas where you want to apply to
fish intelligence and which are there is
we just use plain logic which are the
areas where you try to come up with
things which are very very simple maybe
in some cases sometimes you can call
this cheating but I mean let’s call this
it’s a bit there’s a better word for
this let’s call this hacking and the
main conclusion which we which we drew
from this that apart from everything
else it’s not just mathematicians not
just programmers who need to get
involved with these things these are
people just like you sitting in here
doing many very many different things
being problem owners and having the
understanding on how these things are
built how they function put together
with the intelligence and engineers and
maybe sort of developers can come up
with the best solution for the thing
which not always is just artificial
intelligence which in many cases is a
combination of logic and other things
thank you [Music]
[Applause]
so what you see here is the luggage
built in an airport it’s usually the
place where after a long flight you
expect your luggage to arrive if you are
lucky right actually it turns out in
airport serving around five million
people per year around four to five of
these people decide to take a joyride
on the luggage belt this is actually a
nightmare for security they want to
detect it they want to prevent it they
want to find a way to be able to solve
this issue and we usually are the guys
who are involved to try to solve these
issues and similar like these I work in
the field as mentioned before of machine
vision and artificial intelligence so we
have a team of people being involved in
trying to solve a real-time issues which
might be relevant to different kind of
customers or anybody else
let me tell you how it all started so it
all started back in 2011 we were invited
by railroad company to solve a problem
which was actually solved already long
time before so what they had was a very
simple issue every year around million
wagons in a bit less number of trains
arrives to a country what happens then
as they stopped in the first station
five train inspectors using
walkie-talkies jump out of the building
next with train station and they walk
next to the wagons and they say hey this
is number one this is a regular number
here these are the defects these are
things we find about it this is how it
works it was not really very efficient
process it had a multitude of problems
as you can imagine first of all it is
never a good idea to put the person next
to a 60-ton moving wagon as you can
imagine so that’s a safety issue
second of all it’s prone to human error
so people can make mistakes and you know
when they do so
walkie talkies when somebody writes it
down sends it over fax machines which is
really long time one historical
technology back to the center somebody
reenter this this is not efficient at
all and of course another thing is there
this itself is not really a profession
where you can expect really high growth
but you can’t really become the best in
it or do it the fastest it doesn’t make
much sense as you can imagine so we
weren’t invited to solve this issue back
then and after half a year or a year so
to speak it turned out to be first
failed miserably it turned out we
couldn’t solve the issue straight away
with the technology which was employed
in many places around the world let me
give you some of the reasons first of
all you get yellow numbers and white and
these kind of things
second of all sometimes I get the
feeling people who actually write these
numbers have a special hatred for using
templates as you can imagine there next
thing the vagon to the right it’s
supposed to be clean as you can see it’s
quite a problem to see the number over
there even with your eyes even with the
naked eye that’s another issue which you
might have so we were fighting with this
for a long time then in one of the
meetings one of the researchers who was
not the software developer who is not an
engineer said hey you know but when you
look at the wagons for specific types of
wagons the numbers are in very specific
locations so why don’t we use the first
identification of the type of the wagon
and then we know where to search for the
number which makes our job much easier
so this was actually quite successful it
turned out this way we can do this we
can actually help the guys to solve the
problem so after this we started
actually bragging about ourselves you
know we had this major success that
coincided with the time when artificial
intelligence and machine vision actually
became something like a little bit of a
hype around the world so everybody was
saying this is really cool we should use
this technology everywhere so we were
wandering around the saying yes we did
it we are really good at it we can help
you solve any problems whatever you have
it didn’t take long after a short period
of time we actually were invited by
subtlety to solve a problem they had so
they had two roads one of the road was
leading to a nice beach which in summer
people went sunbathing doing everything
else second one was actually the one
which was more practical that was the
one which was going to the next village
so people were using their bicycles
riding through the shops to sit the town
centre doing all these things and that
was much more practical so they had they
had a question which one should we
develop we said yeah of course we can
solve it that should be very easy right
so the solution we came up with at the
time being was very very simple we said
okay what we can do we can actually see
if there is a motion so we’re gonna
place a camera on a pole we’re gonna see
if there is a motion there if there is a
motion there we will look at the picture
and see if there is a bicycle on the
picture right if there is a bicycle we
say plus one everybody’s happy we know
the number of cyclists sure then we run
into this by some accident somebody
actually drew a bicycle on the pavement
in front of the camera so as soon as
somebody walked by that you know motion
bicycle plus one this was rather easy to
solve right so we trained the system to
understand that this is not an actual
bicycle right so actual cyclist looks a
little bit different and there is a
context in all the other things but then
we run into another problem so we sold
this one and then on a hot summer day we
suddenly noticed there’s a number if
there is a rise in the number of
cyclists but on the wrong road on the
one leading to the next town not to the
beach why so we started investigating
this issue and it turned out there was a
small problem with this as well one of
the cyclists had bought an ice cream in
the local town store and he was going
back home so he stopped somewhere the
best place he figured out to stop was
right in the front of our camera we
didn’t like the guy too much at that
point of time as you can imagine so what
he did he started unpacking his ice
cream movement cyclists plus one then he
he raised it up to his mouth faded
movement cyclist plus-one and within 15
minutes he did it you can imagine we get
like 300 cyclists out of nowhere out of
the blue so you can imagine there is a
special type of as we understood that
the special type of love for the
cyclists versus ice cream and all other
things this was something actually one
of our customers told us how to solve in
one of the meetings he said hey but you
know why don’t you just follow the guy
in the video you have the video stream
right
you just follow him as he moves around
you just found him once this was the
solution we came up with the one we
actually employ which turned out to be
much more successful so it didn’t get
interrupted by somebody eating ice cream
you know pushing buttons in the phone
and all the other things but not all the
stories we had were the ones with happy
endings so let me tell you a little bit
more about the ones we did not so first
of all but start with the one I started
with him very beginning so people taking
Joyride on luggage belt right there were
two ways to solve this the way we saw it
first of all we tried the one where we
said okay let’s train the system to
understand what a human or luggage boat
looks like how do we do it we don’t have
enough data it’s only four people per
year in the luggage about right so we
did we got our guys on a luggage belt in
different kind of positions as you can
imagine and try to create some data to
train the system okay
well didn’t turn out to be that great we
didn’t have enough data still and there
was still some errors so he came up with
another idea actually the customer came
up with another idea he said hey why
don’t we use the thermal imaging where
we can see the temperature of the person
even intoxicated person has a
temperature of 36.6 degrees more or less
give or take it turned out the issue was
that it’s not only that within the
within the all the two days of the
summer we usually get here real summer
suitcases turn out to have the same
temperature because they stand out next
to the airplane you know in the Sun so
that was an actual failure
so that and being really not happy about
having too many false positives or false
negatives this was a problem which we
didn’t manage to solve so this is still
an open challenge which we have then
another one we talked about that one I
don’t think you can see what is here the
the thing on the right is actually one
of the last images one of our systems
took in the small town of Louisville not
so small but still where somebody
decided that it’s well there is a pole
there is a camera why don’t you just
pull it down right so this is no machine
vision this is no artificial
intelligence there was a very simple
solution you can see on the right
you put these sticky things on the pole
and hopefully nobody else climbs it
anymore so that’s fairly simple but
there was another thing related
artificial intelligence and machine
vision which we actually had to solve
and which we didn’t manage the soul so
the customer we had they said ok now you
count cyclists you count pedestrians
everybody’s really happy about it why
just why don’t you count the stray dogs
how many stray dogs are there how many
you know dogs running around without
people you know guy like this really
nice yeah it’s not here mm but still
it’s a bit it’s date time the question
which turned out to be a problem was
that it’s not that easy to understand to
identify even by a person what stray dog
is there’s no human insight it doesn’t
mean the human is not nearby right so
that’s another issue which we have so we
were not really able to identify the
clear definition of the problem what a
stray animal means so how long do you
follow it how long do you see if it’s a
stray animal not so this is still an
open challenge which we still have but
along this way as I said we get invited
to solve different kind of problems
using vision and intelligence but there
were several frustrations which we ended
up noticing along the way and several
learnings which we got along the way
first of all as you can see this has
been a bumpy road so some solutions are
there some are not first one we got was
actually quite interesting whenever you
look at the research when we look at the
data which is presented public
you usually get things like saying you
know this is a sample data we used these
are the pictures and then you just train
the system to say to classify them
that’s fairly simple usually the way it
works it turned out when you have to do
that position of the images yourself you
learn a lot more about what are the
issues with actual photography I’m no
photographer I mean my wife always said
that the easiest way to ruin a picture
is to give the camera to me right this
is actually a clear example of that as
you can see I was trying to catch a car
not that successful yeah but still I and
our team know way more than we would
like to about photography I’m still
crappy photographer don’t ask me to do
any photos but now we know what much
more about what the motion blu-rays and
other things which are related to this
how to build the lighting system and
everything else are these kind of things
but that’s one more there other thing
which is important was research or
frustration many times in our team we
got young guys coming back and saying
hey you know we tried to use this
research which is publicly available
which is the newest thing in the field
everybody should be really happy about
it
but it doesn’t work and then we started
investigating why it turns out that
quite a few problems in that one
first of all in many cases the research
has some small hidden things so it’s
publicly available but some of the data
is hidden second of all in many cases
it’s trained to be used under very
certain very specific circumstances so
that’s another issue which which you
might have right
the third thing which we learned certain
frustration we had was the frustration
of the problem owner
so the problem owner usually was
presented with an idea that you know
artificial intelligence machine vision
is here to solve all your issues right
now it’s already here it’s beating chess
masters in chess your masters and go
learn to play games out of different
kind of things but when you try to apply
this to a very very simple real-world
problems you either end up having
something which is very very expensive
and high-tech or which doesn’t work the
way you intended to work so the
conclusions we had the things we came up
with in the end of the day
actually we’re fairly simple first of
all we understood that it
we need to deconstruct the problem we
need to split it up in pieces and this
is not necessarily where software
engineers and engineers in general
prevail sometimes it has to be people
who actually might have a different
point of view and things people who are
involved with these problems on daily
basis through deconstruct it into small
pieces and then you figure out which are
the areas where you want to apply to
fish intelligence and which are there is
we just use plain logic which are the
areas where you try to come up with
things which are very very simple maybe
in some cases sometimes you can call
this cheating but I mean let’s call this
it’s a bit there’s a better word for
this let’s call this hacking and the
main conclusion which we which we drew
from this that apart from everything
else it’s not just mathematicians not
just programmers who need to get
involved with these things these are
people just like you sitting in here
doing many very many different things
being problem owners and having the
understanding on how these things are
built how they function put together
with the intelligence and engineers and
maybe sort of developers can come up
with the best solution for the thing
which not always is just artificial
intelligence which in many cases is a
combination of logic and other things
thank you [Music]
Please follow and like us: