Hacker News Reader

Comment by dogma1138 | Hacker News Reader

Examples from the company website if anyone is interested:

https://snapshot.parabon-nanolabs.com/examples (TV/Media volunteers)

https://snapshot.parabon-nanolabs.com/posters (Actual cases from the looks of it)

These seems to be for the most part on the same level as some good police sketches, what surprised me is that the few renders where they showed a tooth gap that matched the photo quite well, the seem also to be pretty good at predicting some facial features like the nose but overall fail quite poorly at eyebrows (i'm not sure if it's because of the neutral expression or the algo itself).

Dystopian indeed but if these examples are actually representative I'm almost astounded by the resutls.

View on HN

Surely the company website is only going to put up photos where their predictions closely match reality. You don't get to see all their failures.

While I don’t disagree quite a few of their predictions didn’t match that well so it’s hard to tell how much selection bias there is because we don’t know how many uses they have, the posters seem to be all the cases they worked that allowed the photos to be released, some of the media ones weren’t that close either.

Marketing department smart enough to make the advertisong material pique the interest of the target market, rather than have them, and you, think "they've obviously selected the best possible results".

Considering the method (and generously assuming these are not cherry-picked) I would call these pretty impressive. However, it seems that for practical purposes dispersion is large enough for it to be not that useful. I mean, if you have a reasonable set of suspects, this would help immensely, but basically to a degree that identifying race/ethnicity, age & sex would. Otherwise, it's a pretty generic GTA 3 character model, it would be hard to look for "somebody like that", when you know that the nose/face shape could turn out to be quite different.

Those don't look close at all. If I knew the actual person and saw the predicted picture I wouldn't make a connection.

This is why police sketches don't work well on people with close relationship to the subject of the sketch but work rather well for law enforcement and the general public.

When you rely on a sketch you don't rely on it being a photocopy of the subject but on specific features matching the actual appearance of the subject.

For example eye distance, brow height, nose shape, chin shape matching up 2-3 sketch features with a suspected target tend to lead to fairly high likelihood of a correct match.

Do police sketches actually work well? I doubt it. No one has ever done a controlled experiment to measure the rate of cases closed with or without police sketches.

Looks like they can't predict the age though? Eg. first two examples are way off and say "Age: unknown".

I thought we might be able to do it with the telomeres or something?

IIRC there are sequencing challenges for any section of DNA made up of repetitive sequences. Telomeres are made up of repetitive DNA.

But there is probably some way to do it, there might be age-related "clocks" in the metagenomics as well.

Not everyone is born with the same length of telomeres and not everyone ages at the same rate so it’s not a good predictor unless you have multiple samples over a sufficiently long enough period.

> what surprised me is that the few renders where they showed a tooth gap that matched the photo quite well

So seeing this I had a thought and went back to look. It looks like the renders either have a tooth gap or a closed mouth (and in one case the actual photo had a closed mouth) - could their model just add a tooth gap regardless, or did I just miss one without?

Why did it seem like everyone of Asian decent was predicted to be bald?

It kinda seems like it's just good at identifying the colour of people skin?

I guess that's all the cops need though. It can output "black man" and the cops can go lynch somebody and call it a job well done

If you’re narrowing down a pool of suspects being able to clear everyone who isn’t a specific ethnicity immediately is still pretty powerful.

You don't need to draw a picture of them for that though.

No but if that also narrows down the pool of people further with sufficient degree of accuracy it can arguably be a useful tool.

The only question here is does this actually work in reality and does it provide sufficient benefits without tramping on civil liberties and rights.

Reply on news.ycombinator.com