Geometric Face Recognition - Computerphile

[@The_savvy_Lynx]

I like those computer vision videos a lot

[@superbanaan8175]

Great video, not a lot of people can explain these complicated algorithms in a human way!

[@watchmefail314]

Fantastic sentence to end the video on. :)

[@firebird9594]

or if you have a cat face 😸

[@Steamrick]

+hamena314 Was thinking the same thing. Now there's something to quote sometime in the future. "We all smile the same way." -Dr. Valstar

[@cooldude23407]

or if you were in a horrific accident :P

[@RobvandenBerg290569]

pearl of wisdom at the end : "and we all smile in the same way" made me smile

[@MarioThaMonkey]

This channel needs more viewers. I'm not in love with the content but its very beat to learn how certain everyday things function. Or how certain problems are solved.

[@GaminGeek1811]

Nice video, as always. Hey, I know it's not really your style to talk about games (except maybe Tetris and Portal), but have you guys heard of "No Man's Sky". I think it would make a great topic to talk about that game and procedurally generated worlds, either on this channel or on Numberphile. Anyways, keep on, whatever the topic, your videos are always fun to watch :)

[@I77AGIC]

awesome. make some more videos in the category of neural networks, convolutional networks, machine learning, deep learning, etc. please!

[@MrrGNB]

i have a question concering the feasbility of facial features in certain locations, how does the machine deal with poeple being upside down

[@IMAXXHEW]

+zocke1r if the machine can not properly identify a person, it will set off a flag that goes into danger mode... it will shoot a laser that immediately vaporizes the unidentifiable object. In the future, if the machine can not positively identify someone... that will be unacceptable to Big Brother.

[@ilmansonic]

+MAXXHEW Or it just rotates the picture if it doesn't find anything....

[@thenorup]

+zocke1r Such a shame, I can only give one thumbs up :(

[@IMAXXHEW]

Ilman how many times will the machine rotate the photo to identify the object... how long before the laser is used? We can not predict exactly how AI will work... people believe it will be smarter than us, they say it will learn to make it's own decisions, and what it does could be very unpredictable.

[@ilmansonic]

+MAXXHEW Through binary search it will calculate at which direction it should rotate. Once it finds the correct angle, the layers are used in all angles already checked. That's why we shouldn't give a rotated picture to an AI. :D

[@DrPhoerrets]

I don't know if I'm being stupid so I apologise if it's a silly question. But my understanding is you are looking for a particular position of a feature. You then have an input of pixel coordinates (so some point in a picture). You then put that into a learned regressor (either via deep learning or some other machine learning algorithm) and it comes out with a vector that points to the most likely direction the feature you are looking for is. This vector is learned from all the training data manually set by people. My question is how does the computer decide what each pixel coordinate relates to in the training date? To elaborate, in the video, Dr. Michel Valstar used the eyebrow as an example and said that because it's an eyebrow, the algorithm gives an output vector that points to the most likely direction of the feature (inner eye). How does it know that the pixel value there is an eyebrow in both training and test data?

[@Shangori]

I recognize that accent anywhere.. A dutchy!

[@Huvada]

Yes sir!

[@RobvandenBerg290569]

Dutchies unite!

[@ruinenlust_]

Heuuyyy

[@U014B]

If there's two things I hate in this world, it's people who are intolerant of other people's cultures… and the Dutch.

[@rjfaber1991]

***** You'd be surprised how often those two overlap...

[@nick.raptis]

A video about "eigenfaces" would be a nice continuation. I wonder if they are conflicting or complimentary techniques.

[@Multihuntr0]

I think he got a little bit wrong. He said that Gradient Descent is used to determine if a sliding window has detected an eye and is related to deciding that an eye has been found. I don't know about the exact algorithm they're using for deciding when to stop, but Gradient Descent is a training algorithm. The local optima problem for Gradient Descent has only to do with TRAINING the algorithm that takes a patch of the image says "here is an eye", and nothing whatsoever to do with deciding when to stop scanning through a picture. Otherwise; a good video, keep it up.

[@RenaissancManEng]

My father, Woody Bledsoe, PhD, pioneered the first Facial Recognition techniques in the 1960's at his company Panoramic Research, near Sunnyvale CA, while working on contracts for the CIA/NSA. My dad was a pioneer in the fields of AI, Pattern Recognition, Automatic Theorem Proving and Computer Reasoning, and past president of IJCAI. Many people now use variants of his initial algorithms in many Facial Recognition software systems. It would be nice if he was to receive more credit for these innovative and groundbreaking techniques, IMO. -LW Bledsoe, Austin TX

[@architgupta6037]

I was reading an article about facial recognition algorithms today, and it was righfully mentioned that Dr. Woody Bledsoe was the first one. However most of his work was not published because he was developing them for government.

[@HighSpeedNoDrag]

U.S. Patent(s)?

[@Seegalgalguntijak]

I'd really like to see a tutorial on which changes one would have to do to a picture of a face (i.e. in Photoshop or GIMP), so that it can't be used in order to create a biometric facial profile of a person any more, while still keeping the picture recognizeable for a human, so that a human that sees the picture can still identify the person, but a computer can not. Is that even possible? If so, what has to be altered in a picture for that?

[@elfboi523]

Solution: Use makeup to paint additional eyes, mouths, etc. on your face. You might look like some weird alien from some SF horror flick, but face recognition software will not get any meaningful results from you.

[@basteagui]

nah bro. see, i can map the brightness value of the image of your face to tell me what area has high glossy reflection highlights, and this will tell me where your real eye is as opposed to the ones you painted on.

[@hazzsin3579]

won't the water in your eye also have glossy reflection? Besides to teach the neural network to deal with fake makeup you would need thousands of examples to train it correctly, so it could try to use specular reflections to find which eye is real. And you wont really find thousands of distinct examples easily.

[@ElectricChaplain]

So how is Valstar's group going to scale from a thousand faces to a million faces? Will they take their existing set of faces, find eigenfaces or some sort of basis between vector maps for facial features, and then decompose any new face into a combination of these bases?

[@SteinGauslaaStrindhaug]

Do we have any idea of how our brain does facial recognition, and how it compares to our computer vision algorithms? I guess our visual processing is quite brute force since we have massive parallel processing ability (while computers are better at sequential processing), with dedicated hardware (since brain damages to certain areas can destroy the ability; prosopagnosia) and some kind of probability based heuristic since we see faces in literally anything (pareidolia).

[@ValleysOfRain]

+Stein Gauslaa Strindhaug one of the things you need to get rid of is the idea that the brain is a computer. It sort of is in, the way it deals with information, but a computer bases everything of of calculation and numerical values, whereas with brains it's a little more abstract. If you were to tap into the data feed from a CPU and compare to a similar data feed from a brain (however you did this), you would probably see that the CPU data is deterministic, whilst the Brain data would be wildly abstract. It's like comparing machine code to arabic.

[@ValleysOfRain]

***** AI will follow an exponential growth curve once it figures out it can self-modify. As its intelligence grows, it will be able to modify itself even further, to the point that we can no longer understand what it is thinking. Initially, AI will be slow to grow, but it will accelerate as time goes on. In biological systems, a Natural Intelligence (us) does not have the ability to self modify beyond a certain extent.

[@ValleysOfRain]

Evidently there will be physical limits, but answer this for me: If you were capable of modifying your mind and you knew it would make you more intelligent, would you do it?

[@goodmoneymaker]

What feature type do you guys use when training the classifier ,HAAR or LBP? also which one of the two would you say is faster when using positive samples bigger then 5000 images?

[@learnmarketlevels]

sir am i right in saying geometrical features based algorithms uses less features when compared to appearance based methods for face expression recognition ??? i badly need to understand this.

[@sehwankim9312]

That was awesome

[@TheDebare]

Could it recognize the difference between twins?

[@cgdmtrn]

+Allen Law I assume you were asking for identical twins. It is hard -or almost impossible- even for the human beings to recognize identical twins just from their face. What reveals the identical twins is mostly the way they behave rather than just appearance (or sometimes the hair style, outfit preferences etc but still not only face or facial muscle movements.). So probably I would say no it couldn't recognize the difference between identical twins unless it is deeply trained to distinguish the differences in the patterns of two people.

[@EnricoPiazza]

+Cigdem TURAN In fact you may recognize their expressions rather than their face geometry (A smiles all the time, B is sad all the time. If face is smiley, it's probably A). Would two identical twins smile in the same exact way? Maybe two identical twins have two discernible sets of facial expressions. (Just guessing, I don't know anything about twins or how emotions are translated to facial expressions)

[@GtaRockt]

+Enrico Piazza are there REALLY "identical" twins, like in "their facial features are 100% the same"? If the answer is no, an extremely good software, paired with extremely high quality photos maybe could make out the differences.

[@cgdmtrn]

Enrico Piazza I do not know if there is any physiological or computational studies on expression differences of twins but worth to check it. As you said, we may end up finding a unique way of expressing emotions.

[@jonathanstudentkit]

the errors don't just magically cancel themselves they would have to be biased symmetrically around zero

[@SupLuiKir]

Is it plausible to have a general learning AI algorithm to create its own initial training data, instead of having to do so manually, like you did for the face recognition?

[@AntonioBarba_TheKaneB]

+· 0xFFF1 there are some preliminary studies on general AI and automatic training based on pure observation and inferential rules, but this is pretty much science fiction (yet).

[@2Cerealbox]

+· 0xFFF1 Yes and automated feature detection is almost certainly something a lot of these algorithms are going to use. But sometimes it's easier to just tell the thing what eyes look like rather than hope it figures it out on its own.

[@SupLuiKir]

Ethan Poole Then just teach an AI to teach AIs like humans teach AIs. :P

[@etofigh]

+· 0xFFF1 Yes but who will teach us how to teach an AI to teach AIs like humans teach AIs?

[@SupLuiKir]

Ehsan Tofigh We'll figure it out, probably.

[@bakerbakerbaker305]

This reminds me of those suits they use capture human movement for video game characters

[@casperTheBird]

Why are the errors in opposite directions?

[@SamuelHorler]

How difficult would it be to adapt this technology to CCTV and suveillance, with faces being presented at awkward angles, not directly front-on, slightly concealed etc?

[@IMAXXHEW]

+Sam Horler Do you really think this technology is not being used in surveillance already?

[@SamuelHorler]

+MAXXHEW I have no idea. I'm not involved in surveillance or machine learning.

[@2Cerealbox]

+Sam Horler This is in use, through Bayes Theorem plays a role here. Say you have an algorithm that correctly detects if a person is or is not a wanted criminal 99.9% of the time. You put that to work in a CCTV camera in a public street and you get a bunch of warnings about criminals. Unfortunately, since the vast majority of people are not criminals you're just going to wind up with a big list of false positives.

[@ethanpoole3443]

Often such systems rely on motion tracking and will patiently wait for that opportunity to grab one really good frame to perform recognition on, all the while continuing the motion tracking. Though many will have some degree of capability to compensate for a limited range of off axis recognition so long as sufficient features can be identified (which also allows them to deal with imperfections like glasses/sunglasses or facial hair). A lot of those details have to be inferred, though, as the vendor's jealously guard any intellectual property that they feel will give them an edge on the competition.

[@alexanderreynolds9705]

+Sam Horler This method is robust to faces not directly head on. It's part of the reason this method works so well, in fact. If you are looking for predefined sizes, then it won't work. But for example if you detect what you think is an eye, and it is rotated a bit, then it will point to where the nose should be given the rotation, etc. So this network of self referencing is actually what makes this algorithm work well. Of course, for very extreme angles it probably wouldn't be great, especially on a low-res camera.

[@austinb797]

I still don't think I'll ever get that facial expression minigame in pokemon-amie to work right.

[@tidemover]

Why not use the vertical symmetry of the face? By going to the lowest point of the face the chin. Then Through the curvature of the jawline you find the center of the jaw which is the centerline of the face. Then the mouth is above . Nose is above and the eyes are above and out on both sides of the nose.

[@alexanderreynolds9705]

+Dave Preston Not a bad thought, and some amount of symmetry is likely used - however, not all faces are going to be looking straight at the cameras. Like in this video, he was never facing the camera right on and so if the face is a bit rotated in 3D space you'll have a hard time using symmetry. The left eye and right eye could be different distances from the lens and thus bigger or smaller, etc.

[@alexandrugheorghe5610]

+Alexander Reynolds Also to take into account they are using deep learning algorithms which I assume they take any picture as input the algorithm will sort out which are the markers before even going to detect the facial features for the eyes etc. So if you would want to go with the chin you would first start to compute where it is, normally you can do that but I think it just consumes even more resources than simply learning a few vectors which tell you the feature you're after. But maybe I'm wrong...

[@tidemover]

+Alexandru Gheorghe Oh ok . I just figured that the chin is always the lowest point of the face and if they turn their face the curvature of the chin would be different on one side which could be calculated. I figured the positions of the eyes, nose, ears and mouth are some relative constant based on structural distance. If you found one you could calculate the others centers based on a normalizing algorithm.

[@rentacowisgoogle]

+Dave Preston I like your idea but if i can play the devils advocate, for example, could you teach a computer to differentiate between a narrow face, to one that is slightly turned from the camera? Probably, but how feasible is that w/ modern tech?

[@tidemover]

+rentacow That's a good question. Even damage would be a problem but since the face has a slope or angle . You could look for the symmetric slope and if it doesn't match you could look to see if the other slope is closer to vertical then search for the nose outer point. Then you could work a scheme to get results.

[@chrisbaldwin1156]

Could it handle a close up picture of someone's face while doing a hand stand?

[@IMAXXHEW]

+Chris Baldwin it can't do a handstand.

[@TheDiscoNarwhal]

+MAXXHEW Nailed it.

[@Vykk_Draygo]

+Chris Baldwin The positions are the same, just rotated. It be possible to check for that. I mean, it could just check the points relative to each other, not relative to the photo's orientation.

[@henrypostulart]

OK, so you start every video with a pair of pseudo HTML tags, i.e. then end the video by closing the inner tag ()… I have not yet observed the closing wrapper tag . Am I blind, or do you not understand how HTML works?

[@Computerphile]

please see the video(s) on XML and the concept of tags which do not have to be closed :) >Sean

[@henrypostulart]

Yeah, I've been coding HTML/XML/xHTML/etc. since the 90's, and in that piece of the universe, a self-closing tag ends with />...

[@TheJebusChris]

Is it true you can deceive facial recognition with LED glasses?

[@ValleysOfRain]

+Jebus Chris supposedly IR LEDs create a "halo" that blots out your features to cameras

[@marcelo55869]

+Jebus Chris You could pretty much deceive it with a photo of the person you want to detect. One time I could deceive a cell phone detector with a smiley face drawn on a paper... Nowadays systems can only detect faces.. recognizing different faces apart from one another is a much more difficult task... even the same person can be completely different depending of the situations...

[@JGunlimited]

3:43 But how does it know it's looking at an eyebrow and not a beard?

[@DanFrederiksen]

that regression sounds a little questionable. How about starting with head location in a low res version, then you have narrowed down eye locations quite a bit

[@rasjnda]

This seems like a really dirty technique for finding eyes: algorithms based on a large, manually defined data set. Can't argue with results, though!

[@alexanderreynolds9705]

+Luke Stutters Kinda how machine learning / AI works. You need a lot of data to parse, but once you have that you can make a lot of inferences automatically with statistics and thus starting from a bit of data yourself, the machine learning algorithm could parse through thousands more of images and create a much larger database to use.

[@unvergebeneid]

+Luke Stutters Why am I having a deja vu with your comment?

[@spakwi]

OMG i though i was clicking on a CaptainSauce video

[@jonlanghoff]

Martin Høgsted..?

[@RekMone]

I wonder whether FACS plays a role in this?

[@cgdmtrn]

+Ich Nichtdu FACS is the system which names the facial muscle movements which can also be used for facial expression recognition; and mostly be used. So if you find the facial points more accurately, you can observe their movements more accurately and it eventually ends up a better insight of FACS. And when he says the mouth corners will be wider and upper, it is for sure he is talking about the FACS system. AU #12 - Lip Corner Puller to be exact.

[@RekMone]

Cigdem TURAN I too thought (and still think) it's very likely to be employed here as a tool for analyzing and categorizing of facial expressions. But I guess you're not 100% sure either?

[@cgdmtrn]

Ich Nichtdu I am not sure if they have already used it but I am sure it can be used (:

[@robosergTV]

too bad neural networks work 20% better, so standard CV is obsolete.

[@galesx95]

that accent! it's almost the same as profesor Moriarty! Have anyone noticed this?

[@goeiecool9999]

+galesx95 The accent is Dutch. I think Moriarty is English though.

[@galesx95]

goeiecool9999 missed it by this " much xD

[@alexanderreynolds9705]

+goeiecool9999 Moriarty is Scottish!

[@goeiecool9999]

***** At least I was right about him being a native speaker.

[@knaperstekt7953]

+Alexander Reynolds He's Irish. Don't let them know we can't distinguish between the Englishmen, Scottish and Irish. It won't end well.

[@DrRChandra]

Eyebrows are above eyeballs...unless maybe you're talking about kids hanging off a jungle gym, or somebody trying to stretch out their back by being in inversion boots, or.... :-) It just depends.

[@Eric_D_6]

now rotate the face

[@apburner1]

Computer scientist assisting government in controlling the sheep, yay!

[@saeidyazdani]

I hear some Dutch accent

[@HighSpeedNoDrag]

More so Swede, Finish.

[@kevind.shabahang]

i don't like it...too hand-coded...come up with an algorithm that learns to identify the key features automatically and then we're talkin....i doubt the brain does it your way

[@ValleysOfRain]

+Kevin Durden mmm... don't bring brains into this. They are a completely different beast. The pattern recognition systems in the brain use associations and employ a vast parallel processing matrix. The brain employs techniques for facial recognition which are almost wholly alien to what computers do. The fact is that the brain doesn't quite work mathematically as such. While computers use binary, in the same way that the brain does (eh, kinda) brains also use various neurotransmitters and hormones to appropriate information and disseminate it between neurons. It's a highly complex field, but I can safely say that what works for a brain would not work for a computer. It's a completely different system, with completely separate architecture

[@kirstinetermansen7234]

University Orterd You totally understand the rules of, laws of Commercial abuse, iligal Einterteinmemt, medier Without. Data, truthful. It's worthless

[@jake17124]

First!

[@ariebrons7976]

+Fae Tality firsty? have a drink

[@plk173]

+Fae Tality maybe he's firstea

[@Rottensteam]

+arie brons Mike Tyson pls

[@kirstinetermansen7234]

Also..... Tom cruise. Mask, surgery, botix It's pretty.... Not sure,,,, that. People are twin lol like

[@aqshafeiTV]

This is what microsoft was doing, with there stupid website.