How computers learn to recognize objects instantly | Joseph Redmon

Prevoditelj: Ivan Nekić Recezent: Sanda L

Ten years ago,

Prije deset godina,

computer vision researchers thought that getting a computer

istraživači računalnog vida mislili su da je naučiti računalo

to tell the difference between a cat and a dog

kako razlikovati između mačke i psa

would be almost impossible,

gotovo nemoguće,

even with the significant advance in the state of artificial intelligence.

čak i uz značajan napredak u razvoju umjetne inteligencije.

Now we can do it at a level greater than 99 percent accuracy.

Sad to možemo učiniti s više od 99 posto točnosti.

This is called image classification --

To se zove klasifikacija slike -

give it an image, put a label to that image --

dati sliku, staviti oznaku na sliku -

and computers know thousands of other categories as well.

a računala znaju i tisuće drugih kategorija.

I'm a graduate student at the University of Washington,

Ja sam postdiplomac na Sveučilištu u Washingtonu

and I work on a project called Darknet,

i radim na projektu pod nazivom Darknet,

which is a neural network framework

što je neuronska mrežna struktura

for training and testing computer vision models.

za obuku i testiranje modela računalnog vida.

So let's just see what Darknet thinks

Pa pogledajmo što Darknet misli

of this image that we have.

o ovoj slici koju imamo.

When we run our classifier

Kad smo pokrenuti naš klasifikator

on this image,

na ovoj slici,

we see we don't just get a prediction of dog or cat,

ne dobivamo samo predviđanja je li to pas ili mačka,

we actually get specific breed predictions.

nego čak i određena predviđanja pasmine.

That's the level of granularity we have now.

To je razina zrnatosti koju imamo sada.

And it's correct.

I to je točno.

My dog is in fact a malamute.

Moj pas je doista malamut.

So we've made amazing strides in image classification,

Napravili smo nevjerojatne pomake u klasifikaciji slike,

but what happens when we run our classifier

ali što se događa kad pokrenemo klasifikator

on an image that looks like this?

na sliku koja izgleda ovako?

Well ...

Dobro ...

We see that the classifier comes back with a pretty similar prediction.

Vidimo da je klasifikator vraća uz prilično slična predviđanja.

And it's correct, there is a malamute in the image,

I to je točno, na slici je malamut,

but just given this label, we don't actually know that much

ali samo s tom oznakom ne znamo mnogo

about what's going on in the image.

o tome što se događa na slici.

We need something more powerful.

Trebamo nešto snažnije.

I work on a problem called object detection,

Radim na problemu koji se zove otkrivanje objekta,

where we look at an image and try to find all of the objects,

gdje gledamo sliku i pokušavamo pronaći sve objekte,

put bounding boxes around them

staviti okvire oko njih

and say what those objects are.

i reći ono što ti predmeti su.

So here's what happens when we run a detector on this image.

Evo što se događa kad pokrenemo detektor na ovoj slici.

Now, with this kind of result,

Ovakvom vrstom rezultata

we can do a lot more with our computer vision algorithms.

možemo napraviti puno više s algoritmima računalnog vida.

We see that it knows that there's a cat and a dog.

Vidimo da zna da su tu mačka i pas.

It knows their relative locations,

Zna njihove relativne položaje,

their size.

njihovu veličinu.

It may even know some extra information.

Čak može znati neke dodatne informacije.

There's a book sitting in the background.

U pozadini je knjiga.

And if you want to build a system on top of computer vision,

Ako želite izgraditi sustav na osnovi računalnog vida,

say a self-driving vehicle or a robotic system,

recimo autonomno vozilo ili robotski sustav,

this is the kind of information that you want.

ovo je vrsta informacija koje želite.

You want something so that you can interact with the physical world.

Želite nešto da možete komunicirati s fizičkim svijetom.

Now, when I started working on object detection,

Kad sam počeo raditi na prepoznavanju objekata,

it took 20 seconds to process a single image.

trebalo je 20 sekundi za obradu jedne slike.

And to get a feel for why speed is so important in this domain,

A kako biste dobili osjećaj zašto je brzina ovdje tako važna,

here's an example of an object detector

evo primjera detektora objekta

that takes two seconds to process an image.

koji treba dvije sekunde za obradu slike.

So this is 10 times faster

Dakle ovo je 10 puta brže

than the 20-seconds-per-image detector,

od detektora kojem treba 20 sekundi po slici,

and you can see that by the time it makes predictions,

i možete vidjeti da se za vrijeme dok on učini predviđanja,

the entire state of the world has changed,

promijenilo čitavo stanje u svijetu,

and this wouldn't be very useful

i to ne bi bilo vrlo korisno

for an application.

za neku primjenu.

If we speed this up by another factor of 10,

Ako ovo gore ubrzamo još jednom za faktor 10,

this is a detector running at five frames per second.

to je detektor koji radi na pet sličica u sekundi.

This is a lot better,

To je puno bolje,

but for example,

ali, na primjer,

if there's any significant movement,

ako postoji bilo kakav značajan pokret,

I wouldn't want a system like this driving my car.

ne bih želio da sustav poput ovog vozi moj auto.

This is our detection system running in real time on my laptop.

Ovo je naš sustav otkrivanja u realnom vremenu na mom laptopu.

So it smoothly tracks me as I move around the frame,

Glatko me prati kako se krećem kroz kadar,

and it's robust to a wide variety of changes in size,

i otporan je na razne promjene veličine,

pose,

položaja,

forward, backward.

naprijed, natrag.

This is great.

Ovo je super.

This is what we really need

To je ono što stvarno trebamo

if we're going to build systems on top of computer vision.

ako ćemo graditi sustave na osnovi računalnog vida,

(Applause)

(Pljesak)

So in just a few years,

U samo nekoliko godina

we've gone from 20 seconds per image

došli smo od 20 sekundi po slici

to 20 milliseconds per image, a thousand times faster.

do 20 milisekundi po slici, tisuću puta brže.

How did we get there?

Kako smo došli dovde?

Well, in the past, object detection systems

Nekada su sustavi za otkrivanje predmeta

would take an image like this

uzimali sliku poput ove

and split it into a bunch of regions

i podijelili je na hrpu područja

and then run a classifier on each of these regions,

i zatim pokrenuli klasifikator na svakom od tih područja.

and high scores for that classifier

Visoki rezultati za taj klasifikator

would be considered detections in the image.

smatrali su se detekcijom u slici.

But this involved running a classifier thousands of times over an image,

No, to je značilo rad klasifikatora tisuće puta na slici,

thousands of neural network evaluations to produce detection.

tisuće procjena neuronskih mreža kako bi dobili detekciju.

Instead, we trained a single network to do all of detection for us.

Umjesto toga smo naučili jednu mrežu da učini sve detekcije za nas.

It produces all of the bounding boxes and class probabilities simultaneously.

Ona istodobno proizvodi sve okvire i klase vjerojatnosti.

With our system, instead of looking at an image thousands of times

S našim sustavom, umjesto da gledate sliku tisuće puta

to produce detection,

kako bi postigao detekciju,

you only look once,

gledate samo jednom,

and that's why we call it the YOLO method of object detection.

zato ga zovemo YOLO metoda za detekciju objekta.

So with this speed, we're not just limited to images;

Dakle, ovom brzinom nismo ograničeni samo na slike;

we can process video in real time.

možemo obraditi video u realnom vremenu.

And now, instead of just seeing that cat and dog,

Sad, umjesto da samo vidimo mačku i psa,

we can see them move around and interact with each other.

vidimo kako se kreću i međusobno komuniciraju.

This is a detector that we trained

To je detektor koji smo obučili

on 80 different classes

na 80 različitih klasa

in Microsoft's COCO dataset.

u Microsoftovoj zbirci podataka COCO.

It has all sorts of things like spoon and fork, bowl,

Ona ima svašta, poput žlice i vilice, zdjele,

common objects like that.

obične predmete poput tih.

It has a variety of more exotic things:

Ima raznih egzotičnijih stvari:

animals, cars, zebras, giraffes.

životinje, automobili, zebre, žirafe.

And now we're going to do something fun.

A sada idemo učiniti nešto zabavno.

We're just going to go out into the audience

Samo ćemo otići u publiku

and see what kind of things we can detect.

i vidjeti što možemo otkriti.

Does anyone want a stuffed animal?

Želi li tko plišanu životinju?

There are some teddy bears out there.

Tamo ima nekih medvjedića.

And we can turn down our threshold for detection a little bit,

Možemo malo smanjiti prag detekcije,

so we can find more of you guys out in the audience.

tako da možemo naći više vas u publici.

Let's see if we can get these stop signs.

Da vidimo možemo li dobiti ove znakove STOP.

We find some backpacks.

Nalazimo neke ruksake.

Let's just zoom in a little bit.

Zumirajmo samo malo.

And this is great.

I to je super.

And all of the processing is happening in real time

Sva obrada se događa u stvarnom vremenu

on the laptop.

na laptopu.

And it's important to remember

I to je važno zapamtiti

that this is a general purpose object detection system,

da je ovo sustav za detekciju objekta opće namjene,

so we can train this for any image domain.

možemo ga trenirati za bilo koju domenu.

The same code that we use

Isti kod koji koristimo

to find stop signs or pedestrians,

za pronaći znakove STOP ili pješake,

bicycles in a self-driving vehicle,

bicikle u autonomnim vozilima,

can be used to find cancer cells

može se koristiti kako bi pronašli stanice raka

in a tissue biopsy.

u biopsiji tkiva.

And there are researchers around the globe already using this technology

A znanstvenici diljem svijeta već koriste ovu tehnologiju

for advances in things like medicine, robotics.

za napredak u medicini, robotici.

This morning, I read a paper

Jutros sam pročitao članak

where they were taking a census of animals in Nairobi National Park

o popisu životinja u Nacionalnom parku Nairobi

with YOLO as part of this detection system.

koristeći YOLO u sustavu detekcije.

And that's because Darknet is open source

A to je zato što je Darknet open source,

and in the public domain, free for anyone to use.

u javnoj domeni, besplatan svakomu za korištenje.

(Applause)

(Pljesak)

But we wanted to make detection even more accessible and usable,

No, željeli smo napraviti detekciju još dostupnijom i korisnijom

so through a combination of model optimization,

pa smo kombinacijom optimizacije modela,

network binarization and approximation,

binarizacije mreže i aproksimacije

we actually have object detection running on a phone.

dobili detekciju objekata koja radi na mobitelu.

(Applause)

(Pljesak)

And I'm really excited because now we have a pretty powerful solution

A ja sam stvarno uzbuđen jer sada imamo moćno rješenje

to this low-level computer vision problem,

problema računalnog vida na osnovnoj razini,

and anyone can take it and build something with it.

i svatko ga može uzeti i graditi nešto njime.

So now the rest is up to all of you

Sad je sve ostalo do vas

and people around the world with access to this software,

i ljudi diljem svijeta s pristupom tom softveru,

and I can't wait to see what people will build with this technology.

jedva čekam vidjeti što će ljudi učiniti s ovom tehnologijom.

Thank you.

Hvala vam.

(Applause)

(Pljesak)