SearchFace developer about the capabilities of the algorithm

    Hello everyone, I am one of the developers of the SearchFace service and am ready to talk about it in the comments. Because of the hype with the suit of VC , the important thing for which we launched the service has receded into the background - to test the search capabilities. And since now the service is available to the general public, I want to show everyone what our recognition algorithms are capable of.





    At the moment, SearchFace is a small demo of the algorithms. Each search is carried out in our database of half a billion alternatives. That is, each person must be distinguished from hundreds of millions of others, among whom there may be (and probably will be) people very similar to the desired person. Yes, this task has already been solved by the closed FindFace (if it doesn’t change my memory, the base volume was about the same), so we wanted to not just repeat, but wanted to surpass. The main task that we set for ourselves was to try to make it possible to search, including by highly “distorted” pictures. A few examples below, but you can play around yourself.

    Example 1. Maxim Cherkasov, trashbox.ru.

    Maxim was one of those who, while reviewing our algorithm, did not hesitate to upload photos in mirrored sunglasses. Still, the first three places in the search results were extremely correct results. Moreover, in one of them the photograph was of low resolution, with an unusual facial expression and everything else taken six years ago. Combo!


    Result:



    Example 2. Ilya Dyer and Sultan Suleymanov from Meduza.io. Ilya uploaded his photo, in which he looks away, and Sultan - a photo in a scarf (where only part of the face is visible). According to him, which we naturally didn’t check, he couldn’t identify him on this photo, while we have a very high score on both of the results, which means that the algorithm didn’t just choose the most similar person, but was sure that found the right one. The verge of "confidence" is somewhere around 0.65–0.67.



    Example 3Nikita Likhachev, tj. Revision TJ tested the engine on its employees, but unlike Maxim Cherkasov, she didn’t try to cheat our algorithm. Therefore, for the purposes of this article, we deliberately blurred a photo of Nikita with imagemagick using a gaussian-blur with different sigma values.

    convert Nikita_00.png -gaussian-blur 12x4 Nikita_04.png

    Original photo and with blur

    And right up to sigma = 16 in the first place there was a real photo of Nikita, with sigma = 18 Nikita's photos still came across in the “top 16”, and only starting from sigma = 20 “top 16” no longer contain anything relevant to the request .



    And here is an example of the search for a photo with an unusual facial expressions and unusual angles - photos became infamous this summer model Natalia:


    Photos from the site vklybe.tv

    SERP:



    Here is the result of work on the picture 8-year-old Amy Winehouse:


    In the "top 16" there are a lot of her adult photos:



    With the 7-year-old Madonna, it turned out to be more difficult, but nevertheless, her extradition contains her adult photos, and with a bright make-up.


    Result:



    And here you need to understand that for seven years she was back in 1965, and photographs of those times are not always of high quality.



    It may seem that the photos in the examples are somehow specially selected, but you can try it yourself on any photos. We'll tell about the algorithms themselves a bit later, but for now just want the collective mind to test - fidbek is very important for us.

    Write questions in the comments.

    Also popular now: