undefined | Better HN

0 pointskilotaras8y ago0 comments

> After the faces are located, they’re matched against a set of candidates (which I would guess is an enumeration if all your Nth order friends [who don’t opt out of this], to reduce computational complexity.

Face recognition is usually done by first transforming the face into vector space and than looking for the closest vector from known, e.g. openface[1] (which is based on google 2015 paper "FaceNet: A Unified Embedding for Face Recognition and Clustering"). Facebook itself opensourced[2] a library for efficient search of nearest neighbors in such spaces.

It doesn't seem likely that generating set of Nth order friends (which grow very quickly with N) would be better than just having a global set or a split of global set by some characteristics (e.g. gender, race).

[1] https://cmusatyalab.github.io/openface/ [2] https://github.com/facebookresearch/faiss

0 comments

3 comments · 1 top-level

wpietri8y ago· 2 in thread

Wouldn't the betterness of using friend networks being reduced false positives? Facebook has 2 billion MAU, which means they have a lot of faces. Filtering down to the 50k or so that are close to you means it's much easier to find the right person, rather than a similar stranger.

sangnoir8y ago

> Wouldn't the betterness of using friend networks being reduced false positives?

They are not limiting to friend networks, otherwise they would not catch (out of network) strangers using your pics. One of the use cases touted is "Help protect you from strangers using your photo"

tekknik8y ago

Also finding images you’re in but may have been taken by someone not on your friends list (think events). Eventually probably recommending friends from these images.

j / k navigate · click thread line to collapse