▲ 130 points
back
32 comments
Quite interesting however the portraits really have almost no resemblance to the subjects.
Yeah they look kinda "similar" but virtually every facial feature other than base racial feature is incorrect. All of the renderings attempt to pretty much render a racially correct "grey face" which can pretty much fit almost anyone.
Everything else from the shape and position of the eyes, nose bridge, cheekbones, lips etc. doesn't match the real photographs.
I really hope that this doesn't reach law enforcement because those sketches will fit every "average" person in their respective racial group, and it even seem to get the ethnic group wrong in at least 1 out of the 3 renderings...
The images themeselves aren't that great for identification. However, if they combine it with traditional facial recognition techniques - such as distance between the eyes (a genetically determined characteristic) - it would (for example) be able to narrow down suspect lists to a few people using DMV photos. Wanted fugitives living under false names that have managed to get a real driver's license are regularly caught these days using this technique.
So instead of generating images, this software could simply take a DNA sample and spit out the characteristics the facial recognition software already takes as input. Up comes a list of names and addresses of the handful of people that the DNA can possibly belong to, even if they have never been in a DNA database. The software already has to calculate these things before delivering the facial image, so delivering the characteristics would actually be a subset of the capabilities he has already built. Scary stuff.
Agreed. It seems like nutrition, illness/disease, and random developmental quirks would all drastically affect the actual shape of a human's face between birth and adulthood. You might be able to predict a range of faces to account for these differences, I guess, but then you have a broad range of faces that might match a broad range of people...
WHAT? No! Have you forgotten the fact that identical twins actually look...well, identical? Face features clearly have a very very very very high heritability.
Twins share the same environment in the 1st 9 months, when their faces are formed.
Usually yes but that doesn't change the fact that face features are highly hereditary. I find it very hard to believe that someone could try to deny this. It's basic biology really.
What are you suggesting here exactly? What exactly is the environmental trigger in the womb that forms the faces if not genes?
May I kindly remind you about the concept of dizygotic twins, that is, twins that share the womb but do not share an identical genome. Please explain why they don't look identical.
I'm not saying faces aren't determined by hereditary factors, only that there are also other factors. Almost all of the 'identical' twins I've seen had subtle differences.
Yet those differences are, as you said, subtle, that is, very small. That means that, in the context of this article, there is no huge range of faces we can determine from genes, but the genes (when we know them) actually determine the face very very precisely as proved by identical-twin-resemblance. The predicted faces in the article are fuzzy not because environment has a great impact but because our knowledge of the genes that form our faces is still very limited.
If you disagree, please explain your reasoning, don't just downvote. Thanks.
>I really hope that this doesn't reach law enforcement ...
This was linked in the article itself and at the end:
http://www.nytimes.com/2015/02/24/science/building-face-and-...
This happened to me, an artist took the DNA from my github and created my face back in 2013. Cool stuff.
Do you have a photo of your face in roughly the same pose for comparison? Also, why does the nose look so faceted?
Imagine people with stunted growth or disfigurement as a result of non-genetic circumstance, using a more accurate form of this technology to see what they would have looked like otherwise.
Profoundly sad to think about.
On the other hand, one possible use for such renderings could be to guide reconstructive medical procedures.
Environment plays a large role in facial appearance. Weight, diet, habitual emotional expressions, facial hair, to say nothing of acne, injuries, facial exercises (for the muscles in the face) or reconstructive surgery.
Maybe they could divine how close the eyes are or how big the nose is... but even that doesn't seem terribly accurate from these picture. I suppose the technique will get better soon enough, but I doubt anything approaching publicly recognizable accuracy will be achieved by this method alone.
The inaccuracy can't be solely attributed to environment. Many identical twins who share identical DNA that are separated at birth still share remarkably similar facial features. If two faces generated by identical strands of DNA in the real world remain identical, despite being grown in different environments then an accurate simulation of DNA should reproduce an accurate face regardless of environment.
The inaccuracy arrives from the black box statistical process they use to generate the face. They don't know the full story behind how billions of strands of DNA can scaffold a physical face made out of billions of cells, nerves and blood vessels.
Instead they just measure a bunch of faces relate it back to certain known attributes in the DNA and then generate a face. This process will always be inaccurate. In short they treat the system like a black box and build a simulation of the black box by only establishing correlations between input and output.
Imagine if some alien race used this type of process in attempt to analyze the linux operating system running on a computer. They measure the electrical output of each pixel in the monitor and they find statistical correlations between pixel configurations in the monitor and the arrangement of bits in the cpu register. Using this process the Aliens can only gain a very flawed understanding of what's actually happening.
There isn't really room for anything between this method and outright cloning, is there?
Silly me, I thought everybody's face was DNA-generated...
Those are the results with a training set of 7000, I imagine it would get spookily accurate with a 50K training set.
Probably not. There's diminishing returns with higher samples. If your extra 43K samples follow the same general pattern (without adding some outliers that will give new patterns) it won't look any different. You'd have to capture some gene that gives a difference in appearance that the 7K didn't.
Where'd you get that number?
Just a guess. The faces were pretty close but for a few angles here and there.
I couldn't figure out what a training set of any size would even mean. Taking X number of DNA sequences paired with real photographs, and then extrapolating to DNA sequences without a photograph? That might do something interesting if someone tried it, but I don't think it's that kind of technique at all.
There's another story about the technology's accuracy and its application in law enforcement, here:
http://www.nytimes.com/2015/02/24/science/building-face-and-...
I've been waiting to see something like this, it's very neat (if somewhat unrefined).
Now apply OldBooth and we'll be able to see what our babies will look like. Brave New World here we come!
I think it will need another 5 years of development before we can get even near that.
I wanted to do something similar last year, but I figured that I'd need a lot of input data to classify on facial features. Where do people get the input data?
The article notes that they used data from 23andme, a commercial DNA profiling service that is $100.
My gut reaction to this was so what we've all seen DNA generated faces our whole lives... and nobody finds that in any way remarkable...
The old way took tens of years and a couple hundred thousands of dollars in food and board. Plus there is no way in nature to take a DNA sample left at a crime scene and turn that into a rendering of a face.
Now it can be done with $100 worth of DNA analysis and a few minutes of compute time.