**The World of Face Recognition: A Journey with Linus**

I'm so excited to share my experience with face recognition technology, also known as CompreFace, with you all. As I was setting up this system, I couldn't help but feel a sense of wonder and awe at the capabilities of AI. The more I worked with it, the more questions I had, and that's exactly what we're going to explore today.

**The Challenge of Training**

One of the biggest challenges I faced while training the model was finding high-quality images of people with beards. It seemed like no matter how many pictures I collected, the system just couldn't seem to get it right. Linus, a regular contributor to our channel, offered some sage advice: "I think like once you train a good model with CompreFace, if it only accepted as that person, unless the confidence was like 99%, some of these other models made with people's photos with them wearing a mask." His words of wisdom were music to my ears.

**The Linus Effect**

As I continued to work with the system, I noticed something strange. It seemed to be developing its own personality, and that personality was eerily reminiscent of Linus, our beloved channel mascot. He would sing little ditties about face recognition, like "Oh, wow. Yeah, yeah, yeah." It was as if he had become the embodiment of the system's learning process. I couldn't help but chuckle at the absurdity of it all.

**Security and Implications**

As I delved deeper into the world of face recognition, I began to wonder about its security implications. Fortunately, since we have all of our own hardware and software, none of the data goes outside of my house, which gives me peace of mind. But I couldn't shake off the feeling that there might be more to this technology than meets the eye.

**The Power of Edge TPU**

One question that kept me up at night was: who's actually behind these face recognition systems? Are they worried about AI taking over? Fortunately, our team has been working on some exciting projects, including a server that will be running all of this stuff. We're talking about using PCI Express expansion to create a custom solution, but I'll save that for another video.

**FreshBooks and Our Channel**

As we continue to explore the world of face recognition, we can't forget about our sponsor, FreshBooks. Their simple accounting solution is designed specifically with small business owners in mind, featuring built-in automation that allows you to spend less time tracking projects and more time growing your business. Whether you're a tradesperson, creative agency, or YouTuber, FreshBooks has got you covered.

**The Future of Face Recognition**

As we look to the future, one thing is clear: face recognition technology is going to continue to evolve at an exponential rate. With Edge TPUs and AI-powered systems like CompreFace leading the charge, it's exciting to think about what's possible. Will we see widespread adoption in the near future? Only time will tell.

**The Watercooler**

And that's not all, folks! We're going to be doing an external PCIe box to expand our server's capabilities. It's going to be a wild ride, and I invite you all to join us on this journey into the world of face recognition. Thanks for watching, and don't forget to get subscribed for more content from our channel.

"WEBVTTKind: captionsLanguage: en- This might look like a GPU.Smell like a GPU.Sticking in a PCI Express16x slot like a GPUand even talk like a GPU.(muttering gibberish)But I assure you,its intended purpose is very differentand to some potentially terrifying.It's dubbed the ASUS IOTAI Accelerator PCIe Cardor CRL-G116U-P3DF for short.And it's designed for artificialintelligence computation.Speaking of intelligence,you'd be intelligent to knowI'm going to tell youabout our sponsor Honey.Honey is the free to use shopping toolthat helps search for the best promo codeson tons of your favorite sites.Get it today at joinhoney.com/LTT(upbeat music)Historically, we've mostly shied awayfrom covering AI stuff.And there's a coupleof good reasons for it.One is we're not AI ormachine learning developers.So finding practical use casesthat we are able to set upand demo for you guyscan be a little trickyand second, a lot of those usecases are extremely technicalbehind the scenes kinds ofthings, at least for now.So hardware like this,isn't too applicable toend users like me and you,that isn't till today.While doing research for and testingthe home automation setup at my new house,get subscribed by the way,you don't want to miss any future contentaround revamping that place.We stumbled into a bit of a roadblockregarding presence detectionor the ability for the systemsin the house to be aware of whether or notanyone is actually home.Presence detection providesus with a number of benefits.For example, we can turn off lightsand set the HVAC to bemore energy efficientwhen there's nobody at home.And this is relatively simple to implementon a whole house level,but therein lies the challenge.We put a lot of effort intomaking my house a lot moregranularly controllable,practically every room has its ownindependent HVAC and lighting,which is awesome for personal preferences.Like the kids mightwant their rooms coolerthan my wife and I do forexample or vice versa,but it can also be usedto improve efficiencyand save costs.If no one's in the giantrec room downstairs,what's the point ofheating and cooling it.I mean surely that can'tbe that hard to automate.Well as we found out, itcan be really freaking hard.One idea we had was to installBluetooth low energy beaconsin every room and read people's presencethrough their phones.The only problem with that isthat people have this tendencywhen they're at home to puttheir phone on a chargerand oh, I don't know,walk around their house.So that won't work.Another idea was to use motion sensors,but those kind of crapped the bedas soon as you decide totake a nap on the couch,and then you wake up in asauna of a room in the summer,which sounds like a pretty bad time to me.So this is where our AIcard comes into play.This is an ASUS designed product,but all of the important AI bitsare actually made by Google of all peopleunder the brand Coral.We've got both of them linkeddown in the description below.Let's take this thing apart.This specific card has16 onboard Edge TPUsas Google calls them.And the TPU stands forTensor Processing Unit.In the most simple terms,these TPU's are hardware processorsthat are physically designed and optimizedto run a specific application.You may have heard ofsomething called an ASICas something that's really good at mininga specific cryptocurrency.Well, these are the same idea except thatinstead of mining Bitcointhey are for runningMachine Learning Inference Calculations.So what we're lookingat here is essentiallyjust a PCI Express 16x interface heregoing you can actually see the tracesgoing right into this PCI Express switch,which appears to be justsplitting out these lanesinto an M.2 2x interfacefor each one of theselittle Dual TPU's here,I was expecting something lesskind of modular andscience fair project-y.Now each of the Edge TPUof which there are twoon these M.2 cards can doaround 4 trillion operationsper second with a powerdraw of only 2 watts.So 2 watts times 16 TPU is brings usto 32 watts for the card.You add in your PCIe switch and fansand that brings you to 52 watts,which is well within the 75 wattsthat this PCI Express slot could power.But for some reason it still requiresa six pin power connector.And it's got this giganticcopper fin cooler on it.Now you're probably thinking,wow, that's really cool Linus.But most of this sounds likethe boring technical stuffthat you were talking about earlier.But if you were following closely,while we were researchingbetter ways to detect presencein my new house,we stumbled upon somesoftware called Frigate,that uses cameras plus this hardwareto really efficiently detect presence.It's really freaking cool.Let's take it for a test drive.I got to put this back together.And we're at this test bench is runningthe latest version of unRAID in itand it's going to standin for my home nest.Now you might notice there isalso a GPU installed in it.So we need that as a video outputbecause this is a non-APU Ryzen Processor,but it's also useful for offloadingthe camera video decodeprocess from the CPU.Now in my home deployment,I'll probably just use the CPU becauseI'm going to have 24 epic coresthat are otherwise going to go unused.But for now, this is the better solution.Now, once unRAID's booted up,you can see that we havejust a single array discwith no redundancy.That's not what I'd recommendas a production deployment,but it's good enough for this video.Now we just pop over tothe community apps taband installed apps.So we've got our Coralaccelerator module drivers.We don't need this ifwe're using a USB versionof Google's AI processor,we only need it for thePCI Express version.And then we've alsogot our NVIDIA drivers,which are for our GPU.Now it should be noted thatyou can use your CPU for Frigate,but performance is not good.Next, we install Frigate.Now we can't just launch it.We've got to actually configure Frigate,otherwise it's going tohave no idea what devicesit's supposed to be using.So step one, we're going to copythe GPU UUID from our NVIDIAdriver and paste that there.Then for our TPU mapping,we've mapped it to our Apex_0 device.So that is only one of the 16TPU's that are on that cardbecause we're only going tohave a couple of cameras.We're not going to bothermapping all of them,but we did map five of them.So that should be enoughfor our demo for now.Wow this is a greatpassword, Jake. I like it.♪ Are you serious? ♪Don't get overwhelmed.Most of this is included in the template,but there are a couple ofthings that we had to tweak.So one is we had to set up our camera,so you can see we pulled our camerafrom the writer's denand we've got our networkpassword right here.We're just using the one camera for now.We actually didn't takeall five of our NPUs.We've just got four ofthem configured here,again this is total overkillfor what we're doing.And then we've also got this right here.So we are limiting our frames per secondfor the camera feed thatwe're using to 5 FPS.That is all we're going toneed in 99% of situationsto have functional detectionand adding more processing to thiswould serve no real purpose.One other critical thingback up at the top hereis that you can plugthis into an MQTT server,which is basically itsway of communicating backto your other home automation devices.So this MQTT server 10.20.0.71is running on our officehome assistant instance.Hey, 5 FPS of green.I can see why you saidwe didn't need more FPS.There's not a whole lot of movement there.♪ It's peanut butter demo time. ♪♪ Peanut butter demo time. ♪♪ Peanut butter demo time. ♪♪ Well, yeah. ♪♪ Well, yeah. ♪- Okay, That one camera is set up.We have another one too,but let's just...- Okay, hi there.- Oh, wait, hold on. I screwed up.- Oh! All right.Do we not have the detection on?- We do but I, I didn't turn it on,like, the bounding boxes.- Bounding boxes are justlittle boxes that show uswhat the software's detectingand the degree of certainty that it hasfor what that object is.Humans aren't the onlything that it can detect.But they're kind of the onlything that we care aboutfor home automation presence detection.- You're telling me wedon't have to detecthow many apples youhave in your fruit bowl.Oh yeah. Oh yeah.Oh, you're, you're there buddy.Oh, it's kind of detecting your feet.- Does it know who I am?- That part we have not done yet.- What if It only seeslike this much of me?- Yeah it stillthinks you're there.- What if my legs show?- Oh, it lost you.- It lost me? What about a hand?- Oh yeah, it sees you again.- Oh, really?Okay, what if it's just so an elbow?- Oh yeah, it's detecting you.- What? Really?- Yeah.- Okay. Can it can it see me?- It's confused right now.- It's a little confused?What about if there's just like a casual,like hand there?- It's kind of freaking out.So sometimes it actually sees two peoplewhen he's kind of obscuredbut that's totally fine.- Oh like, it thinks Imight be like two peoplehaving a (beep) behind the curtain.- Yeah sure.But it's fine because all we need to knowis if there's anyone, right?- Right.- Yeah. And (beep) aside.- You don't want the heat tocut out during your (beep).- Okay, all right.- You will get cold.Man! You can do so muchcool stuff with this,but wait, there's more.- What if you have a common area,like a kitchen or a rec roomand you want to be able toapply personalized preferenceswhen you're present, likesay have your Spotify musicwith your specific playlistplay when you're in the roomand not when Yvonne's in the room, right?- Oh, I never even thought of that.- You could walk around the houseand have Spotify playing here.Oh, I'm over here now.And now there's my Spotify over here.- Oh my God. That would be so cool.- Turns out someonehad that exact thoughtand created Double Take,which is a piece of softwarethat allows you to determine the face.This is more creepy.- Yeah.- But it's all local.- I mean we could justconfigure this to not logany of this (beep), right?- Yeah.- Did a Linus appear?- Oh, Wow! Yeah.- Nice.Now Double Take isn't availableon the unRAID community appstore, at least not yet.So we need to manually set upthe docker container ourselves.Ben wants it set up.We need to configure it tolisten to Frigate over MQTT.It should be noted that Double Take itselfdoesn't do facial recognition.It's just an API andinterface to make hooking intoand training facialrecognition software easier.So the first thing we need to dois grab some pictures of me,which are going to be training datafor our facial recognition.So I guess I need to go over there, right?- We could go over thereor we can just uploadsome photos we already have.Double Take currently supports DeepStack,CompreFace and Facebox.They're all machine learning basedfacial recognition models.DeepStack is on the community app storeso that's the easiest to set up,but I tried it out and itwasn't really that greatwith the limited training data that I had.CompreFace, on the other hand,it requires Docker Compose,I have to run into VM,but it's scary good.I put a couple of photos of Linusand secretly kind of had himwalk by without telling him.And it was like, Linus.And then I walked by and it was like, Jakewith like 10 photos of each of us.- All right, let's try it.Why do you have so many picturesof me on your hard drive?(both laughing)- Well you make my drive hard.- Why CPU when I can see you pee?- Why talk to you whenI can leave the zoom?How do I use a Mac, do you know?- No! These stills are horrible.- It's Scary.- Good Lord.- It's you without a beard.(Linus singing)Oh, wow. Yeah, yeah, yeah.- Oh my God, 95.8%?How can it be that sure onunbearded Linus training?- There was only one that itis actually confident about.- I mean 94% seemspretty damn confident to me.- You know what wecan do is take all of theseand say train off of those data.Now you wanna walk over there again?- Yeah, for sure.- Okay. That's plenty.There's so many Linus' now.Linus, Linus, Linus, Linus.Like there's a couple of timesthat it thinks it's James.I think like once you traina good model with CompreFace,I'd probably say that like,if it only accepted as,as that person, unless theconfidence was like 99%,some of these other models,Madison, also made with people's photoswith them wearing a mask.So like it literally has the eyes.Can you go back?Run over there and just...- All right, I'm going,I'm going, I'm going.- See if it makes some more mistakes.- I'm going to pull upa chair, Linus style.- It's veryconfident about Linus now.(Linus struggling)- Who else would behave like this?- I don't think that'swhat it's looking for but...- Man so you could just kind of wanderaround your house for 10 minutes,find anything that'swrong, boom, train it.- Part of it is also goingto be specific camerasbecause you know, each cameralooks a little different.- One thing I askedJake about and he said,it doesn't have thisfunctionality for now is,can it morph over time andcan it continue to self-train?- The thing with facial recognitionis like the more data you feed it,the more uncertain it becomes.- All of this leavesus with some questions.Do you even need this thing?Could a bog standard CPUdo the same calculations?The answer is yes, itcould, but way slower.To put it in perspective,your average Quad Core CPUwould probably only be ableto handle a few frames persecond, on a Quad Core.But because these EdgeTPUs are so optimizedfor running these specific calculations,they are stupid fast at them.A single Coral Edge TPU can handle arounda hundred frames per secondof people detection in Frigateand we've got 16 of them.So if my rough math is correct,that's around 1600 frames per second.We might actually be pushing thatonce we've got all thecameras in the new place.- But we're onlytaking only 5 FPS for me.- Still, it's a lot of cameras.- I don't know 5 FPS.- Yeah, five. Okay,yeah, man that's crazy.So a single one of theseTPUs could probablyhandle 15 to 20 cameras at 5 FPS.Another question this raises is,are we worried about thesecurity implications?Fortunately, since we haveall of our own hardwareand all of this is opensource and locally hosted,none of the data goes outside of my house.So I'm actually pretty happywith the level of security and safety.As for being worried about AI taking over,no one's asking that question.So it's in here anyway.The answer is no!Leaving only one questionthen, who's our sponsor?FreshBooks.FreshBooks is the simpleaccounting solutionthat is designed specificallywith you in mind,the small business owner.It features built-in automationthat allows you to spendless time tracking projectsand more time growing your business.So whether you're a trades person,creative agency or a YouTuber,you can choose the planthat's right for youwith FreshBooks.They have an Award-winningToronto-based support teamwho are always happy tohelp you if you need it.And you don't have to take my word for it.You can try FreshBooksfor free for 30 days todaywith no credit card requiredat freshbooks.com/Linus.If you enjoyed this video,maybe go check out the videowhere we built the serverthat is going to berunning all this stuff.We're going to have toget a little bit creativewith PCI Express expansion on it,but that's the machinethat's going to somehow runthis full AI card.- Yeah, yeah.- So good luck everybody.Are we going to do an external PCIe box?We should totally do it.- Yeah, we should totally do it.- Yeah, that'll be a video.Get subscribed.- We have watercooler and everything.\n"