Working on the [Redacted] project, and I'm really wondering if it's better to run one LARGE face recogniser (ie: the bit of kit that has a neural net/thing that says 'it's XYZ'), or have one per known person, and then run them all over every thumbnail and have the one with the smallest distance (ie: strongest claim of recognition) as the suggestion.
That way, when it guesses wrong, you can say 'ok, ignore THAT person, give me the next best result', which isn't possible with one large encoder.
Faster to learn one person (as you can do quick encodes of each person), which means you can add more to the encoder as you go, rather than batching a single, long encode
Slowwww recognition as it runs the encoded data over X many people, where X is a large number already and going to be astronomical, eventually
Maybe you can counter the last point by deprioritising some folks? Like, say 'prob. not X, Y or Z so don't bother unless I ask.'
@DarkestKale could you have some kind of index, then run over the thumbnail and check which recogs it's worth sending to for proper scanning? I don't know the tech so this may make no sense. Although I have actually edited papers on image recognition algorithms!
@DarkestKale@shimminbeg For me this weekend:
Saturday: get it working in Perl
Sunday: rewrite it in Crystal and add the extra features that would have been dog-slow in Perl
Add comment