 Thank you especially for being here so late. I know, you know, beers are later, so Hopefully this will be quick That's me, Javier Ramirez, and at Technical Evangelist at AWS, that means that I speak to technical audiences about what's possible using cloud services But today I'm going to be speaking not only about what you can do with AWS But also about open source recognition, a bit about the science or the algorithms Behind face recognition, so hopefully it will be interesting and I have a few demos and hopefully they will not break And that should be even better That's me on Twitter, Superacronyne. If demos break, it's a good moment to take pictures of me crying So that's still quite interesting and I'm going to be speaking about Facial recognition of a real-time video with the bonus of serverless, but that's you know, just a different thing. So I Don't know But if you're like me when you were growing up or maybe getting older and not maybe growing up But when you were growing up you saw movies about you know the future for me The future was like you know speaking to my watch or those things that was like super cool But actually in 2002 this was basically the future. I don't know if you've seen this movie maybe anyway, so this is a minority report and Yeah, there's not playing but that's okay So minority report and this was like a very interesting part in which Tom Cruise He was like you know like moving along the city and he was like getting like Customized advertisements you want a beer you want whatever and then he was entering a shop and and they were like hey If you want you can go to the right which is where we have the clothes that you like So that was like super cool It was like real time recommendations with fresh recognition and that was like you know like kind of the future Yeah, and this was like the first time I saw this was like I don't know It looks like a bit fake especially the blink in the eyes was like I don't know but actually This is not a movie anymore. This is a Thank you that's a supermarket in Seattle So you go there and you need you need to be a member So you need to first take a few pictures when you register But you enter the supermarket you go there you grab anything you left the supermarket and when you go you get a message Oh, you bought this and we are going to charge you for that. That's about it So you don't have to you know to stop at the cash register or wait or a scan the items that that's so 20th century so that's that's for real that's you know supermarket that Amazon did the They commerce not Amazon was services were different companies They have in Seattle and that works with face recognition and also with some sensors But of course, it's not only fiction and only Amazon. This is in Korea So in Korea they have like this like you know Korea is very big and they have a lot of people everywhere So basically if you try to go a concert and you need to scan your ticket It's going to take a while something you can do if you buy the ticket in advance you can scan your face and Assign the face to the ticket at home. So when you go to the concert you go through it through the kiosk and they say Oh, I know you are this person and you have this ticket so you can enter and we already registered that this ticket has been used So that's the no that works same thing if you go to some airport That's in China and you go to the airport you pass the know the check-in Security and at that moment they take your picture and then you can go to places like this in which they are going to tell you Oh, cool your flight is waiting on gate e 25 or whichever the gate which is again kind of cool and that's in the wild that's production That's production radius is walking or again China you go there you raise there for the system and You can just buy things without having to show any payment method because you already registered your face And that's good enough. So this is just to show that there are a lot of use cases interesting use cases in production That's in which people are using face recognition. And if you don't know about this, you might be thinking, okay But they might they must be like very smart people that you know, they are doing very cool things as a developer I tend to think they are better than me at searching that is a cover flow because that that's how I do things So if you know what you are doing, it's actually very simple Very simple to have this kind of systems today And you don't really need to be an expert machine learning if you are good for you But you don't really need to be an expert machine learning for this and a lot of use cases interesting use cases in the wild in production today for machine learning and face recognition for example One that to me was very interesting If you have some kind of gambling addiction, you can actually tell the casinos. Hey, you know I like to gamble and so but I have this problem So please don't let me enter your casino because I'm going to spend all my money and it's good for you But for me and you can do this and if you do that, they put you on a list But it's very difficult to enforce that actually now some people are using a face recognition to do this You enter the casino and you know, the cameras are searching and it's not for seeing who is trying to treat the casino It's just for saying oh this person has a problem and she told us that you know, hey, don't let me play So they can and they tell you you remember you told us not to let you play come on come on here here And that's super cool. I mean it's very interesting and then you have like the creepy people like if you are in that, you know dating site you can upload a picture and tell I want people looking like this That's kind of weird, but you know it works So they are face recognition in a lot of places today in production walking. This is not the future anymore This is you know, the things we are living in so if you want to do something like this Hopefully not the dating site, but if you want to do any of these other Use cases, I'm going to tell you, you know, how this is possible So the thing is it sounds kind of cool at least to me that you can do face recognition in real time And serverless but have only three problems first the video processing second the face recognition and then doing the serverless and it's kind of interesting That's what I'm going to be talking about today How you can use different things to actually make a system walking system in which you have real-time video recognition without having to manage any servers And the first issue I'm going to talk about is facial recognition and I don't know you but If someone asked me a few years ago, hey, I want you to do Something to the test faces in a picture. I don't know you but I would have said this is a hard problem And and it's actually hard and if you tell someone which is not in software They might be thinking it's actually very simple. I don't know. That's a face. It's basic human thing Yeah, you see a face, you know who that person is but actually it's if you try to process images It can be as you all know quite Intensive in processing. It's not an easy problem to to work with so how we can do face detection face recognition using software available today. Well, the first thing is differentiating between face detection and face recognition. So face detection is very easy I take a picture and I tell the system Tell me if there's like any human any face in the audience and the system will say yeah approximately I don't know 25 people I was expecting more people But you know like 40 people or something but I don't know who they are but as you know That's the first thing if we are talking about recognition It will be actually telling the system giving these two images or these image database and this picture Tell me if the person of interest is on the database So they are two different problems But in order to solve the second first we need to go for the first one first We need to understand if we have any face in the picture and after that We are going to see if we can recognize who is in the picture You're still with me here. Cool. So the first one face detection is actually simple enough So you want to do face detection? There are a lot of algorithms that have been used through the years to do these kind of things They can be more or less sophisticated But in the end they are quite similar and some of the simple ones actually some of the simplistic ones Very simple. There is one. I will do a demo. It's called hard cascade And in that one basically what we are thinking is wait, wait, wait, wait I know this is hard problem But I already know that the face is usually it's something with two eyes and a nose in the middle and a mouth You know, and that's kind of a face. Maybe not for everybody Some people will be like, but I don't have two eyes. That's okay. That's okay. I get it But for most people, you know, that's that should be good enough. And actually we know that the Saddles around the eyes are a different shade than the rest of the face and so on and so forth So a lot of systems are cascade for example What they do is like a quite simplistic algorithm in which we Train a system to detect those kind of patterns just to detect if we can find Something which is like kind of a square and you know, the status like the different and then something vertical in between And then something else. So something that looks Approximately like no like a face So basically we pixel a day makes quite a lot and we're trying to find the boxes matching that pattern And that's simple enough, but it works. You want to get fancy here. You can do things like no DNS so we can have like deep neural network to detect faces It also works, but you know with simple methods actually finding faces is easy enough. They are then there's some Slightly more sophisticated algorithms and some that are not really sophisticated, but they work kind of the skin color is like Yeah, it's sensitive because you know different people different skin colors I get that but if you train a system with different states of a skin Then look at the proportion of how much skin you can see it can try and the test it is a human in the image So the first or them kind of simple enough the second one. I'll tell you later slightly more difficult So if you want to do a face detection system yourself And if you want to do it from scratch because you know you are in big things a spin So you really like to spend time doing something that has been already done I don't know but if you want to do this from scratch you can actually do it Just take a few thousand strongly several thousand images Normalize them choose the model you want to use for this hard cascade or whatever you want to use train the system and That's it You have a model that it's working and if you want to find Some face data sets for playing there are plenty of them. You know all the mayor players Microsoft Google Facebook a lot of universities they have released Data sets with faces that are some of them are pre-labeled so you can play with them and it's like you know you can Train your systems with that or you can try to do that in a bit easier way. So I'm going to show you the first face detection thing I'm going to be using an open source Library are you familiar with open CV? Open CV is a Python library to work with anything with images not only for face detection can be used for anything else But it's actually simple enough. So I'm going to start that you Peter notebook you Peter Type in is not easy cool Okay, so I Have here one example of face detection using Open CV and that's all the code you need actually we have a lot of code here We don't need that many code for this. So the first thing is you import the library It works and has you know some models already Then I will tell you more about this function in a second and then here we are doing two things I'm going to be loading the model the data of the hard cascade model the one I told you that Models a face based on the distance Between the eyes and the color set around the eyes and the nose and so on so I'm going to be using this model with This this algorithm with two models one the one for the Tatina face in general and another one for a face with glasses because you know I wear glasses and make things slightly different. So I'm going to be loading this So I just point the system to my XML with the with the model coordinates. I Start the webcam This will start and then I'm just calling this function This function has a lot of things the first thing is I need to normalize the image Because when I was training the model when this model was trained the hard cascade It was trained with images that were gray scale So I need to convert to gray scale because otherwise the model is not going to the test very well so I need to first convert to gray scale and Once I have that all I'm doing here is just saying the test as many faces as you can on the On the video and for each face. I'm just going to be Drawing a square around but that's all I'm doing. So here This has started and he's saying oh, it looks like a face. Okay, not really perfect, but But you get the idea. Yeah Oh, this looks like a face. Yeah, kind of I don't know. Maybe the light maybe not and if I point this to my image On the video. It also can you see there? It says like a lot of faces here It's all but but that's the thing so it's it's pretty simple It's like it seems there's a human and it's very easy to treat this system actually very often If I point this to my t-shirt, it's going to be trying to find things where they are known But did you get the idea? Yeah, so these kind of systems are actually very sensitive to things like the lightening different the difference on different lights You know, or if I if I do something like this I'm not a human anymore because it's like, oh, where are the eyes? What is the nose? I don't know So these are these systems are very cool for walking with faces, which are looking, you know forward and with The light which is like no not really many say it's not many contrast, but they walk so open CV is used very often in places like industrial environments not really for first recognition, but for doing quality control using video because in those environments you have like a very controlled set in which you can actually control the lightning you can actually control many things But as you can see if you have the libraries and the models it's pretty simple So this is just with open source and and you can just you know Restart and you can just do that and it walks out of the box nothing to do Okay, we are good to go. Oh, I didn't want to do that. That's okay Hopefully it will disappear Cool. So this is if you want to do just you know something open source simple enough There is another way to do face detection, which is slightly easier, which is just calling an API So if you are using AWS with a service called recognition I'll tell you a bit more about recognition in a moment But with recognition you can actually call an API and say this is my image Tell me if you have if you can find any face and it will tell me things about the face It will tell me things like this is what you have the I don't know the The tip of the nose. This is the left corner of the eye the right corner of the eye This is like, you know the the left part of the of the upper leaf blah blah blah blah And with that you can do interesting things once you have a face and the coordinates You can make a simple application to put I don't know hats or mustache on on top of the people and it's like I don't know that sounds like a pretty stupid application Yeah, that's a snapchat and it makes billions on people just putting hats and things on top of glasses on top of the image So, you know, you can actually do that in a very simple way, but oh, sorry You cannot see it on my screen. I have something saying I don't have any updates to you know to do so That's that's one thing But we are going to do something a slightly a bit better because I told you face detection is easy enough But face recognition is where the things get more interesting. It's like, okay How can I go from I know there is a face somewhere here. Do I know who this person is and Yes, unfortunately right now you still need to tell the the system beforehand the people you want to look after I have a friend here in the audience. He was telling me but if I had to tell you before the people that are in the system That's tricky. Okay, it's not that smart But the thing is how you do face recognition the first thing we do is we find the the face itself and then we take that a square with the face because Applying, you know any kind of algorithm on a small part of the image is much easier than applying to the whole image Second thing is all the algorithms and they are plenty. They try to do the same thing They try to convert that image into some vector that conveys as much information as possible Of a face so the first algorithms I get faces which has been around forever It tries to do this, you know in a in a more traditional way new algorithms like the open face which is based actually on Google algorithms, but it's open source. They do this with neural networks But in the end all of them what they try to do is try to find the things the landmarks of the face the eyebrows The eyes the nose and they put a vector so hopefully if I have my face in one picture and that vector and I have several Images of my face the vector will be always the same because the the relative position and shape of my Main features the landmarks of the face. They are always the same So basically all the algorithms try to do the reduction to a vector and once I have the vector Then I can try to do something interesting the cool thing about neural networks for example open face is that they can actually reduce the The vector for any face to around 128 bytes So with 128 bytes I can have a very small vector to represent any face and that's important because all the algorithms they will need more space to represent the Information of a face and if you have more space and you have ten thousands of faces It's going to be a slower to you know to train and to the test So the cool thing about neural networks since they take into combination or the convolutions They can actually reduce to a smaller shape which is really the innovation that we have seen with neural networks lately Once you have the faces reduce to a vector. We are talking about a classification tolain a very simple classification problem I know this vector is this person this vector is the same person this vector is this other person So once you have the vectors is just classification and you can do classification with anything in this case You can you can actually use Svm You can use actually boost you could use a different network to do classification But once you have the reduction to the vector you can classify in a very simple way that sounds kind of easy enough I don't know and and you could actually try and do that yourself But what I'm going to tell you is like how you can do this without needing to use these models yourself But if you want to open face is open source you can use it you have open face for the model You have the face data sets so you could play with that and you could get good results But I don't have the time for that. I want to do something a bit more interesting So I'm going to do a small demo. Sorry a small demo of Recognition for face detection. Let me ask for a second what I have this Recognition cool So This is the one I told you at the beginning that you can just you know here Just get the different lab marks of one face which is not that interesting But I'm going to go to this one which is the one I want to show you so here with recognition I can upload two pictures. So here I'm going to be uploading Let me see where I have these Here it is. So this will be that's an image of a conference in which I spoke a few months ago and Hopefully when it loads, so he's telling me this girl is not it's not is no one on the other picture That makes sense. So I'm going to upload a picture of myself with the Queen of England I Was getting pretty citizenship. So as you can see there is telling me, okay It looks that in these two pictures this person and this person is the same I 99.4 Sure that they are the same person, which is true. I mean, it's me and we recommend by the way never go below 90 sorry 99% And in some cases don't go below 99.5% depending what you are doing But in this case the system is telling me this seems to be you and you are not this you are not this you are not this And thankfully I'm not that person. So, you know, it's like, okay, it works But I'm going to do something slightly more interesting This is a picture of me and my school friends 37 years ago humans Can you recognize me in this picture? Who you say I am a Machine is about to beat you. Who am I in the picture? Any takers? Even my friends any takers over there? No, and you know me and you know me for years Who is me in the picture? You don't know you don't know. Of course, you don't know I mean, I would have seen this one because it has a bear. So it's like it's too tall I know but for the face it might be me but this is not me. It's not me So, so, yeah, no one cool. That's okay. Let's see if recognition can Can't tell me so if he knows who is me I should have here when I upload the picture a result telling me I am 90 something percent percent sure that this is you 37 years ago. Let me see. Let's see if this is walking Let's see if we have to be scared or not so Come on, come on, say it's me say something. Oh What happened it's like, oh no Yeah, sad yeah, but But let's look at the gays on because this is not a marketing event. This is a technical event So let's look at the gays on and let's see what's happened here. He's telling me first. He's saying Well, I'm pretty sure it was a face in the first picture. That's cool There was a face on the same in the first picture then for the first phase that I I told you you are not that person I was 53 percent sure that you could be that person, but I wasn't sure it was only 53 percent For the second phase is actually telling me solely 28 percent and the third one is 27 percent and yes this this guy here with the dreamy eyes and The petete sweater, that's me. That's me. It's only 58 percent sure But it's much more sure than about the others which is like not too bad Yeah, yeah, so that's kind of the thing 37 years ago and We could argue I look slightly different now. I mean my eyes are closed I don't know wearing glasses. I don't know. Yeah, I need a haircut in both pictures That's that's for real, but you know other than that the similarity is like I don't know So that's what we're talking about you can implement these things actually just with an API call today Which is kind of interesting, but still like okay heavier This can be so it's a nice trick or whatever, but we want to talk about you know Video and about all the things so just before we get there How you work with recognition the first thing you do you create a face library So basically you upload a few pictures of each person and you really don't need much You don't need many pictures So if you want to recognize someone you upload a few pictures two or three should be enough for many cases So you have a few pictures of one individual and you label the pictures this picture is this person This is this person so you put an ID a name Whatever you want to put there and then you ask recognition Can you recognize anyone in this picture and that's how it works But what if we want to do this of a real-time video, which is the The what I wanted to demo today So for real-time video because you know we want to do like a movie like hey, you know You want to track wheel is made going into the underground and The the suspect is I don't know where so you need to have like a lot of video cameras and so on so basically the thing with Real-time video is like you are going to have CCTV several cameras somewhere You need to be able to aggregates the you know the content of the cameras You need to be able to process that somehow if the system is getting a hit You know detecting someone you truly want to be able to replay the video So you can actually because if the system is telling me the suspect is I don't know in with platform I might go to the video and actually check. Oh, no, it's not that person or maybe it is I don't know but you want to replay the video be able to check if something is there and And you want to do that in different devices so trying to do that You know it might be already is there is not a special difficult But you truly don't have the in-house knowledge to work with a streaming video and to do those things And where you store the video how you replay so the good news is there is one service We have which is kinesis video streams, which is specialized on this kind of thing It's a specialized on getting Video feeds from different places if you want to do analytics in this case We're going to apply it to machine learning, but it will be for anything actually it works not only with video It also works with audio and with rather but anything that you have a signal and you want to do Realtonality you can use video streams for that and it works and it's fully serverless So the cool thing about video streams and recognition is that they are very well integrated So on kinesis video streams, I can actually on recognition I can create what we call a processor which points their recognition library Sorry, I was boring. So it points the you know, you know, I know I had a problem. It's not the first time so it points recognition your face library to the video stream that you choose from kinesis and Once you have that, you know, it works automatically. I'm going to have the demo in a bit But before we go there part of this Talk was about doing this serverless because the cool thing. I mean if you are in business today the single most important skill you can have is Producing value. So that's that's it. It's not about. Oh, what's the most important skill? I don't know. I really I know a spark That's cute. I know machine learning. That's cute But you really need to you know to produce value and do things as quickly and as efficiently as possible And that's where serverless, you know comes into so serverless is one of the buzzwords of you know lately And serverless doesn't mean that they are no servers. Of course, they are servers, but you don't have to worry about them so serverless means is Someone else told them is not my problem. So serverless means you can deploy your solution and Since you don't have access to the servers, you don't have to worry about how to scale them The service will be able to scale up and down because it's not your problem I mean the serverless provider doesn't give you access to connect to your machine So they have to you know, somehow manage scalability on their own because you don't you cannot do it Sending for security if it's serverless. It means I cannot connect to apply a security patch So the the provider needs to take care of the ticket. Sorry to take care of that part also serverless also means okay, so if I cannot connect to anything and If he's going to scale automatically if I'm not using the service, I'm not paying anything. That's right That's data from serverless. You don't have usage. You don't pay you have a lot of hits you pay more But that's kind of the thing from serverless. You can actually do you know Walk a bit better and since we are we cannot manage the servers if there is an issue the provider should be able to work with that So the way the city of Amazon puts this is there is no server is here to manage that no server So if I don't have any servers, that's it. I don't have to you know I can focus on other things. So basically you are able to produce things faster You are able to innovate faster You are able to just focus on the things that make your business unique and forget about this daisy things like oh We are getting a lot of hits. We need more processing power. That's cute. We are running out of hard drive We need to are more hard drive. That's okay Those things you don't need to have like someone specifically for that So serverless means that you can actually do this, you know without having to worry and we'll talk about serverless Sometimes people speak about serverless only on the content of serverless functions But serverless can be applied to any part of the stack today We have serverless databases they disconnect when no one is using them They scale to as much as you need when you use them We have serverless data lakes you can do big data on serverless You know things that the kind of things you will do with Hive you can actually do with serverless technologies in which you only pay For that when you are running queries when you are not running any queries You don't pay for anything. It can also be applied to miss excuse. It can be applied to API management It can be applied to any part of the stack So for my particular example the one about video recognition I'm going to be using serverless at many different points first I'm going to be using kinesis video streams for processing the video feed once The video feed is is entering kinesis video streams Recognition will try to see if there is any faces so face detection first and second If any of the faces are from any of the persons I have on the library Then if something if there is any match or if there is like any face even if it's unknown on the feed Recognition will send a message to a message queue and that message queue It's again going to be a serverless message queue in this case is called kinesis data streams Which is something similar to Kafka. So it's just a message queue We are going to have a lambda function which is listening to that queue So if there is an event it's going to do whatever with or with the message and in this case What I want to do is just send a post somewhere So I'm going to be putting that into another notification service, which is going to be calling an HTTG post So a lot of different things just to not manage any servers But with this I'm going to be able to have like a fully managed solution to you know to work with this It was it looks a bit nicer with this So the webcam is going to be my laptop what come from here. I go to kinesis video streams Recognition will be Reading from the feet of the video and I told you real time. I was lying. It's going to be about a five second delay So if I second is real time for you, this is real time if I second is not real time then this is kind of Batch, but this is five seconds delay. So Recognition is going to do that. It's going to be using the face collection. We have on the library From there if there is a match is going to be the message queue will have a function I believe it's in Python might be give us the Bible if he's Python We have Python which is just passing the message and sending it not to a cell phone But to an HTTP endpoint that's kind of the area. So let's let's look at this You have actually the code available if you want to play with that and if you want to deploy It's pretty much work with one click, but I'm going to show you all of this walking. Hopefully So first things first If I go here to my interface, I have something called kinesis video streams Which has a stream here which right now is not assuming anything because I'm not sending any data from the webcam And that's what we have. I told you Let me just stop Jupiter and make this bigger Cool So I told you in the door to work with recognition. We need to have a collection of faces so here you have a connection which is called facial recognition demo and If I Want to see I not typing. Okay, that's okay list Okay, so here I can see these are the faces that I have indexed So basically I can see we I have only three images on my library and the three images are on myself The first image is me and the system is 99 99% sure that it was me on that picture the second picture is So this is the first one. Yeah, and the label is heavier The second one is again me and the label is again heavier and in this case the system was hundred percent sure it was me when I upload the picture and The third picture It's a game have yet hundred percent match I have only three images and with three images I'm going to see if when I move around the video this works and actually I took this in this pictures Long time ago a few months ago and it should work. So I have this and I told you also I need to have something called It was there the stream processors So I have here an extreme pressure which is a stop. So I'm going to start it start Okay, if I describe now The street professor it will tell me this is already Started so now all the infrastructure will be there for me to send the feed what will happen is so I have Kinesis video streams. I have the stream processor if there is a match What will happen I and still not I still didn't start the webcam and I still not sending anything But just for you to see the the whole picture all the pieces if there is any match We have something called Datastream which is similar to Kafka a message queue in which you can see for the last hour I've been doing some tests to make sure this was working So it will be just sending messages to a message queue and for this message queue I have a very simple Python function very simple Python function serverless On serverless functions you pay you pay only for how many milliseconds a function has been executing So here you have a function Which I can see is reading from the Kinesis datastream And this is the code is just very easy Yeah, this actually JavaScript So it's very easy code Everything I'm doing here is like if I'm getting any heat I will be getting a JSON with the content of the heat So I'm just getting the image from the the content from the Jason Sorry, and I have on the Jason two objects the Matchet faces and the faces that I recognize, but I don't know who they are So I'm going to be putting that I'm going to be just counting. I'm going to be I have here Just for each face. I know I'm going to say who is and for each face. I don't know I'm going to see it was not this person. I send this to a message queue and Again, this is a bit more complex that it's all but it's just for you to see that there are many things you can do With serverless and here I can see that this Message queue is going to be sending data to some HTTP endpoint, which is let me just copy this So I have this endpoint here Which right now Is not getting any data I'm going to the history because there was only demos So I have this endpoint and right now we are not sending any data because I didn't start the camera That's kind of the idea but that those are all the pieces. Okay, no servers No, anything just code and things that talk to each other. So now I'm going to start the camera So this is starting and now we start sending data and right now there is no one here So, you know, it's not going to say anything But if I put myself here and if I go to Kinesis video streams again What I have it Kinesis video streams. What are you here? If I go to my stream First thing is I should see here data already The image so okay, so this is the image with a little delay, but that's basically the image So this is what I and of course I could like no go back and forward as you can see here So I can play with a stream and move back and forward and this is just me collecting the video The interesting bit is that if everything goes well, I Should be getting here hits telling me I'm 100% secure this shirt. This is happier and then we can do things like Let me see if this works If I point this to you, let me see for a second. So this is with some delay But hopefully after a bit Actually with a lot of delay now It looks like 30 seconds not five. But anyway, what can I say? But hopefully after a little bit it will catch up. I don't know if it's that oh, yeah and We'll see with the light if it can detect something but I Not sure so let me see if I go to the other top which is tricky to do So if it can detect someone with this light, it should tell me I can see people But I just don't know who they are on the test someone as I said, you know with this in this light and the laptop like that It might not be working pretty well and the thing is if I disconnect from here It's not going to but if someone wants to come here like you know, just to put here and and see that it's actually someone there oops If I put myself here back on the video, it's will say oh, yeah. Oh, what a no person on the video Okay, it's now catching up. So it can already see people. I'm hopefully at some point. It's like, oh have yet again But that's the thing it kind of it can find one and no person it can find Maybe something like that. We'll see more people but that's the idea So you are getting in almost real time like you know the feet from what's happening here Which was what we wanted to get from the presentation and as you could see to do this all we have to do was Starting the feet just sending some you know just writing some Function to a parse the event and do whatever we want with that in my case I'm just putting this on HTTPS point, but you could use it for absolutely anything you wanted as of today. Oh, actually now oh Now we could see like no I could see it to people, you know So it's a bit behind the times, but it's still like you know working there So if you want to do this use this today DSD case available in Java and C++ and also for Android so you can you know play with that a little bit and recognition is available on any You know on any platform, but the streaming part just the streaming part is available only on those platforms and That's That's basically what I had what I wanted to leave you with is if you don't know that this is possible And you can actually Implement that in a very simple way on your application without having to really understand machine learning Without really having to manage any servers You might be thinking that this is the future, but actually you know in the meantime other people are using this already for interesting use cases The second thing is I know this is sensitive You know material I know physical condition. It's kind of you know, so use it for good. I Just this week I saw some news very cool. Actually, I'm going to see if I can find it here I know I'm already over the time, but just for a second. I Saw the other day Face recognition protest Washington Recognition Amazon Let's see if it's there cool. So Man, okay So these people these are privacy privacy concerned people in the US is completely completely legal. Oh, you cannot see Okay, I don't know what's going on. Anyway, I will tell you if you cannot see I'm going to ask to disconnect and connect again Cool. I don't know what's going on there, but I just I just search on the on the internet something It's it's a basic skills. You just Google and you get results. So yeah, so it was oh, thank you Thank you very much. So basically these people are privacy, you know I wear people and in the US is completely legal to be in a public place scanning No recording the faces of people and and these protesters they wanted to raise awareness of this issue So basically they put the self this self this happened last week They put a set this this funds here. They use in recognition the same library I just told you and they feed the system first with Faces of politicians and press from the US and they were in front of the White House Yes, you know just watching people come by and and then if they were getting like any hits the politicians could see on the screen We know who you are so they will be aware that actually this is a sensitive thing So maybe we need some legislation. Maybe not but you know, that's kind of the thing It's actually very simple to do it's it's kind of cool that they were protesting against recognition And they are using recognition for you know for doing the protest So it's you know, it's it's also been proven for activists But I fully agree when you are working with this kind of things you need to be aware that you know There are some limits of the things you should be doing even if technically it's possible That's what I wanted if you have like any feedback or any questions or anything That's my Twitter and I'm going to be around for the rest of the afternoon. So thank you very much for being here