The best face identification system is the ONE that You tested
In high technology, such a high level of lies that you need to check everything. And in this article we give a method of verification available to any person, as well as tell you about all the tricks colleagues in the field of demonstration of facial recognition systems Deceive everyone, we too, so do not believe anyone – all check yourself!
Our system GOALcity different from others is the fact that it is completely free to download and test. It goes to 4 channels of biometrics for free for eternal use, i.e. it can be tested in an unlimited long range of time, revealing all theoretically possible troubles. And moreover, here we will explain how to experience these extreme loads. Naturally, this applies to any system of biometric identification of persons.
A lot of pitfalls: what You show in the demo rooms, often only tricks. In high technology focuses the most high-tech. To be different from our "prophets", we have to release free versions in order for people to appreciate the reality.
For the tests of the biometric identification system, we connect a live camera in front of which the extras will walk – attract several of your friends – at least three, the number in our case does not matter, because we also give the base. Someone from friends, store in database – someone not to identify false positives (triggering on those who are not in the database) and false-negative errors (triggering on those who are in the database, but not on the right one).
On the right we put a panel that shows and highlights all the faces that appear in the camera. And this is also an important indicator of quality: if the system does not find all the faces, it is either due to the fact that it does not have enough resources – it does not have time, or a bad algorithm. On the left – only those persons who are in the database. Here we have to see the correspondence of those people who come to the camera with the names from the database that appear in the identification process. Once again: on the right should appear ALL the persons on the left - only those whom you have entered into the database, and strictly under their names.
The easiest way to throw a face into the database can be directly from the facial recognition system GOAL. A person approaches the camera, a GOAL finds a face in the frame, displays the video of this person on the right panel – and now you can enter all the found photos into the database, assigning a unique name. As you can see, the process is very simple and convenient – you do not need to take pictures of anyone somewhere, upload somewhere, check something there. Well, in the special Lab tried! – it was a second of advertising, conscientious advertising.
First, this person is entered into the General base of the GOAL system. Then we can each panel 'Attention!"set a rule on which persons from the base to react. And so individually for each dashboard with its individual tasks: somewhere you need to show those who steal, somewhere those who buy a lot, somewhere those who need to notify the management personally... Here we are the left panel "Attention!"give the command to bring citizen Hareline when it appears in front of the camera.
In the GOAL system, you can enter not only one photo, but several, or rather a lot. In the future, if, for example, at some terrible lighting, a sharp turn of the head or an updated make-up system does not recognize the brow from the base, you can always add this photo to the floor of a new person – thereby douobuchiv SHOT down. And this is another nice feature of the GOAL system. Oddly enough, most of our competitors have only one field for photography.
By the way, here, if you notice, too, the illumination is not so hot, and the pictures are semi-blurred, but the neural network from the SpecLab pulls them at the expense of its own internal force – we trained it for such "wild" conditions. And this is also a distinctive feature of our neural network. Most of our competitors teach it in pictures from the Internet – and you know what pictures people post there! Of course, everywhere clear oil-free face in good light and successful perspective. And the special Lab has a cloud service with video surveillance records from real objects of hundreds of thousands of customers, and there is a real wildness! And this real life is the basis of the school program for our neural network brought DOWN.
But all the tests known to us are carried out on the combed photos, so we will not deviate from this rule yet, in order to have equal conditions for testing with other systems – we guess that you will test others. Let's start with the classic tests, where we make sure that SHOT down in the special Lab is the best. And once you promise to release others.
And I must say that the most critical in the comparative analysis of systems. How to evaluate them at all? We have been dealing with this topic for 22 years – since the pre-Indian times. On our site you can find a lot of old articles about this, but we did not release anything to the market, and even when competitors have already sold such a function in full. It was funny to look at the enthusiastic customers who claimed that their machine recognized. But the trick was that the computer had no choice but to find out. Of several people it naturally somehow found the right person. With a small number of almost any shot DOWN will work perfectly, but with the base of more than a hundred people often begin to make mistakes. And this is logical – people are very similar creatures.
Therefore, for testing, we give our base for 10 000 people. Most of the world's tests are carried out on this amount. It is easy to download from our website and download through a special file Downloader, which can pick up the entire folder with pictures at once, and at the same time calculate their validity – the presence of persons in them. Why did not we build this database into the program? Only for one purpose – not to deceive you. Then we would know exactly who not to respond to. Already built-in base – it's like a marked card, they are easy to put a mark, so as not to take into account, because in your office there is no one of these people. And here's a clean picture, which, moreover, you may shuffle or prerelativity yourself – here to deceive is difficult. At least, they need to recognize and identify, if system it can, then why bother to cheat?
Naturally, you can upload your database: any images of faces of any size that you can only find, for example, on the Internet. Here again there is a place for advertising, GOAL allows you to quickly and easily download any photo album machine. All files are checked on the fly for the presence of persons in them and are hooked to a database with unique names. For example, so you can in a few movements to download the entire passport database of the country, if you have it.
All persons hovering visible, all the names show up in General, here we have all the civility and solid facilities. To prove once again that we do not cheat with our database, the program has a choice of the level of similarity of persons. With small values of shot DOWN will often give similar to your statistician personalities. And only at levels in 55 – 65% of spoons will be gone. Our scale is designed for very wild conditions, so 65% - is the equivalent of a hundred percent in shot DOWN, trained in "combed" GOST pictures.
Database uploaded, added to even a few extras from people, we start the test. The GOALcity system currently has more than 10 000 persons, and its task is not to embarrass myself with the recognition of your extras, that is no different from the other tens of thousands of entities. We show a camera shot DOWN, capture monitor, and separately removing the Amateur camera.
So, all the faces are left, and those who are in the database are right. On the left we have all who pass in front of the camera. Their appearance on the panel – the fact that they are calculated parameters.
We are pleased to note that only the correct names appear on the right panel. We hope you have, too. If not, specifically for your conditions, your light, your view and so it is necessary to select the accuracy level move the slider to the left to the right.
If one of the participants is missing, then there are two options: either our algorithm is bad, or your computer is very weak. Try using a more modern PC, such as an I9 processor. But here is the percent of the previous generation I7, and it seems to cope.
In General, catching a face is already the solution to half the problem. And to catch at high speed – a virtuoso solution. And how can you achieve high speed? Neural network eats resources without measures. Processing even a single frame takes a long time, and the person at this point can turn away, scratch his head or just be out of focus and in the glare. I. e. it is necessary to stand in front of the camera and wait for a few seconds when the ACS will work. This is unpleasant, especially in the flow of people, where you have another 10 people in a hurry to work. And with a large number of people in the frame – loading resources even more. Therefore, the special Lab uses powerful video cards such as nVidia – the same as on farms for mining bitcoins. But you are testing on a regular PC. To speed up the process, special Lab has developed a technology for the distribution of frames by processor cores. One frame – the first core, the second – the second, and so in turn. You only need to specify in the settings – how many cores you have. For example, here 8. And this is another nice feature from Spetslab, without this multithreading you would not have been able to test everything on a regular computer with a regular Windows 10.
But you need to understand that the number of simultaneously recognizing the faces of the cameras will decrease. Therefore, you have the opportunity to redistribute resources on the cameras. Where the face is held in front of the camera for a long time, for example, in a long corridor, one core is enough to have time to read data from it. Where, as in our test, these are moments, we need to allocate more cores. And it is always possible to scale processors or video cards to increase volumes. For the test, of course, use the maximum number of cores on your PC.