SearchFace Developer on the Capabilities of the Algorithm

Due to the buzz around the lawsuit with VK, the main reason we launched the service—to test the capabilities of our search—has taken a back seat. Now that the service is available to the public, we want to show everyone what our recognition algorithms can do.

Currently, SearchFace is a small demo showcasing our algorithms. Each search is performed against a database of half a billion alternatives. That means every person must be distinguished from hundreds of millions of others, some of whom may look very similar to the person being searched for. Yes, this task was previously tackled by the now-closed FindFace (if I remember correctly, their database was about the same size), so we wanted not just to repeat their success, but to surpass it.

Our main goal was to make it possible to search even with heavily “distorted” images. Below are a few examples, but you can try it yourself.

Example 1: Maxim Cherkasov, trashbox.ru

Maxim was one of those who, while reviewing our algorithm, didn’t hesitate to upload a photo of himself in mirrored sunglasses. Even so, the top three search results were all correct. One of the results was a low-resolution photo, with an unusual facial expression, taken six years ago. Combo!

Example 2: Ilya Krasilshchik and Sultan Suleymanov from Meduza.io

Ilya uploaded a photo where he’s looking to the side, and Sultan uploaded a photo in a scarf (where only part of his face is visible). According to him—which we didn’t verify—Facebook couldn’t recognize him in that photo, but our search returned a very high score for both, indicating that the algorithm not only picked the most similar person, but is confident it found the right one. The “confidence” threshold is around 0.65–0.67.

Example 3: Nikita Likhachyov, TJ

The TJ editorial team tested the engine on their staff, but unlike Maxim Cherkasov, they didn’t try to challenge our algorithm. So, for this article, we intentionally blurred Nikita’s photo using ImageMagick’s gaussian-blur with different sigma values.

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

Original photo and with blur applied.

Up to sigma=16, Nikita’s real photo was still the top result. At sigma=18, his photo still appeared in the “top 16,” and only at sigma=20 did the “top 16” stop containing any relevant results.

Another Example: Searching with Unusual Expressions and Angles

Here’s an example using a photo of the model Natalia, who became scandalously famous last summer, with an unusual facial expression and angle:

Photo from vklybe.tv

Search results:

Result with an 8-Year-Old Amy Winehouse Photo

The “top 16” included many of her adult photos.

Searching with a 7-Year-Old Madonna Photo

This was more difficult, but the results still included her adult photos, even with bright makeup.

Result:

It’s important to note that she was seven years old back in 1965, and photos from that era aren’t always high quality.

It might seem like the photos in these examples were specially selected, but you can try it yourself with any photos. We’ll talk more about the algorithms later, but for now, we just want the collective mind to test it—your feedback is very important to us.

(To provide feedback, follow the “source” link at the bottom to Habr and leave a comment on the article.)

Our other channels