Proof 1
Find yourself in a crowd
Upload a selfie and a photo with several people — a graduation, a party, your team at work. The engine detects every face in the second photo and marks which one is you.
A face recognition engine that indexed 3,995 wedding photos and 15,681 faces — and returns someone's own photos in about three seconds.
Below are two proofs you can run right now with your own images. Nothing is stored.
Try the engineUpload a selfie and a photo with several people — a graduation, a party, your team at work. The engine detects every face in the second photo and marks which one is you.
Two photos of yourself — with and without glasses, front and profile, today and years ago. The engine does not compare pixels: it compares the geometry of the face, as a 512-dimension vector.
The project started from a concrete problem: the photographer delivered the archive as a Dropbox link, and every guest would have to open photo after photo to find their own. The guests' photos are not public — what follows are the numbers.
Each bar is a similarity band, checked face by face by hand. That is how the cutoff landed on 0.40 — below that, accuracy starts to fall.
Above 0.60, every match was correct. Between 0.40 and 0.60 you get side profiles, backlight and one case of sunglasses — all correct on review.
SCRFD locates every face plus five landmarks — eyes, nose and mouth corners. In a group photo it finds dozens in one pass.
Each face is rotated and cropped to 112×112 through a similarity transform, then turned into a 512-dimension vector by ArcFace.
Since the vectors are normalized, similarity is an inner product. Searching 15,681 faces is one matrix multiplication — which is why the answer comes back instantly.