Reminder

Reminder.

Attached: 1567458995606.png (1024x964, 970K)

Other urls found in this thread:

arxiv.org/abs/1908.08239
twitter.com/NSFWRedditImage

>a machine can do the same thing as a human at first glance
Terrifying.
Come back when those faces are properly censored, then you can shit up the board.

FPBP

its not about uncensoring images, its about identifying people by few pixels from a street cam

predicted deformities
what have they been feeding that system?

Attached: face.png (70x152, 15K)

Is there plausible deniability in the predicted faces to say, 'that's not me'?

>identify from a street cam

except these are using staged and posed photo examples, so already have a much more limited variables for the outcome.

look at how bad it fucks up the backgrounds

Some of those would make good "My face - My soul" pics.

it fucks up the faces enough.

What if this is a ruse?
>pixelate some face pics
>mess up some face pics
>violá

Most of those people have two mouths.

Nobody uses 4k street cams.

all this horse shit today in 2019 yet
-i cant properly translate a document from one language to some other language
-any kind of voice to text or viceversa is unusable garbage

but hey thanks at least the computer can recognize a face seen 30 feet away that is very important progress.

I'm failing to see how helping you read Japanese porn is going to help me track the whereabouts of citizens 24/7.

>mfw

Attached: Capture.png (163x168, 78K)

nice, now all those blurry pictures can help find criminals from street cams.

Probably perfectly good data tainted with disableds for maximum western social credit output.

social credit system incoming

how viable is this in videos? (i.e. jav)

You're wrong though.
Only dictatorships use mass surveillance against their citizens, right? That's what they said at school in history classes

it's absolutely not accurate enough for that and will never be admissable
look at how it even misassigns eye colors

And cars and planes are impossible too right? i better get a horse then.

Link to paper/code user?

jav seems a lot better application because the variance is smaller. if will need a good amount of uncensored jav, then censor it, train a GAN and you're done

If you can reduce the search space from absolutely infeasible to something manageable it's totally worth it.

Attached: 1540006285036.png (600x444, 544K)

The only features you can rely on are the ones that a human could already recognize from the pixelated pictures so it's not worth it at all

I call BULLSHIT

they trained this on the same data that they used for testing later
how do i know?
pic 2 row, 2 col: this guy's white collar cannot be predicted from provided 16x16 input

either deliberate fake or total morons

This. Most "ML scientists" do this to hype their research.

>you can be identified from a photo taken of your face
All the cyberpunk literature in the world could not have prepared me for this dystopian future.

If you help user read japanese porn, then you know with some certainty that he'll be reading japanese porn 24/7.

I'll be worried once it can do the same thing with less than ideal lighting.

arxiv.org/abs/1908.08239

Honestly the results look less and less useful the longer I look at them