Love this package. has a bunch of functions for most face recognition, detection and feature extraction purposes. love the readme.md docs for familiarizing with the basic features. PLUS it offers different models(backends) for the tasks which lets you try out a bunch of approach for the same task without writing a custom function
The performance is focused on correctness and the APIs work with individual images. The underlying models can be run with batches, but you need to extract the architecture code to run directly. As a result, while I started with DeepFace, I mostly just used the models.
I'm curious whether others here are working with supervised dimensionality reduction for face embeddings, particularly using single-task or multi-task learning approaches.
While clustering tends to perform better after dimensionality reduction, selecting the optimal dimensions depends heavily on your specific use case. This makes it more complex than simply applying PCA or t-SNE.
If you have labelled data, you can try linear discriminant analysis (LDA; also known as canonical discriminant analysis), which maximizes the between-class variance while minimizing the within-class variance to best separate different classes by projecting data onto a new space that maximizes class separability (for whichever classes help your specific use case).
Very nice package. I used it recently for a project where I needed to detect faces in images as tasks with celery [0]. I wonder if there is an equivalent for OCR.
How do you measure positive and negative societal impact of this technology?
I find mobile phone face unlock so useful, giving every citizen the power to use face recognition could be better than a few people, robots that identify someone and give them lifesaving medication are great (but the opposite, robot assassin can also be created). I guess it comes down to good people building good tools. Humans are generally kind and empathetic
how accurate is the age detection? i.e. lets say you have someone who looks much younger than their age, would this model be able to detect the persons actual age?
For most machine learning systems, "performance equivalent to an expert human" is the best you can expect - simply because it's learned from training data labelled by expert humans.
So if a person looks much younger than their age in the judgement of an expert human, I wouldn't expect the model to do any better than that.
You should also know a lot of work in this area relies on photos of celebrities scraped from the internet - which is much easier than getting loads of labelled images of normal people in normal situations, which would be a total hassle practically and legally.
Of course that has some benefits - if you know the celebrity's date of birth and the date of the photo, you don't need to rely on human labelling to know the age of the person in the photo! But it has the major disadvantage that if your application doesn't involve professionally made up people with movie star looks in evening wear on red carpets - you might find real world performance falls short of the benchmark claims.
> For most machine learning systems, "performance equivalent to an expert human" is the best you can expect - simply because it's learned from training data labelled by expert humans.
However, this is one of the few cases where presumably the data could be perfect, far exceeding the ability of a expert human.
Some things have no ground truth. For example, masking objects for training vision object detection models. Does the object end at this pixel or that one?
It depends on the quality of the image. It's most accurate for a square head-on photo, like a passport photo, and can be wildly off for occluded or angled photos. It's more accurate for people under 30, for older people it often underestimates age.
i would assume it CAN Be accurate if people 'look their age' but of course you can easily have someone who looks younger or older than their age that skews the overall etimate
Quite curious how no one ever talks about “responsible use of facial recognition” or policies to control the use of facial recognition, as it totally pervades and destroys the ability of a person to remain anonymous, at all.
It’s always curious to me how the peasants always seem eager to facilitate the interests of the monarchs to oppress them rather than their own interests to remain free from control by the narcissistic psychopathy of the ruling class prone to tyranny. What do you do as a peasant once you’ve closed the trap you created and led yourself into? I guess maybe more accurately would be to say that it is the aspirational minor nobility that facilitates the creation of the structure that serves the creation of oppressive, top down structures. It’s an odd human characteristic.
Your criticism seems somewhat misplaced. An open-source facial recognition library enables the peasants to wield the same tools that the monarchs already have at their disposal. The cat is out of the bag, and there’s no putting it back.
We can ban the stuff until such time when we truly need it.
Doesn’t help with the climate crisis? No. Does it help with any of the ongoing health threats? No. Does it help hungry people finding food and healthcare or somehow advancing science or any cultural benefits? No.
Does it consume vast amounts of water and electricity to facilitate “bad example of humanity” use cases? Oh yes.
Just feel our priorities are not where they should be.
So we as humanity should stop all actions that are not directly solving the worlds largest problems (at least if there might be slightly negative side effects)?
And what’s up with the energy and water consumption argument. I accept this for LLMs maybe, but not for a local python script that runs on consumer hardware. Should we stop playing video games, too?
Respectfully, I think it's more likely that you're just not personally plugged into and/or have your awareness tuned to them but those conversations are definitely happening. There's not a single consensus (at least in the U.S.), but discussion definitely occurs and in many cases has led to concrete action.
There is the EU AI Act which heavily regulates the use of facial recognition, but I feel that we as “the people in tech who makes these things” should be a lot more conservative in creating frameworks, abstractions and generally advancing/facilitating facial rec use.
It’s even more shocking as this library also incorporates a great deal of cultural bias. e.g. gender, emotion are attributes which vary a lot more than what the models allow for.
Technically there is average IQ by country site, just google. Not that difficult to get faces by country. Put them together. Of course there are regulations and ethic. But in some cases it should work well and is more or less acceptable. Like on Down syndrome or alcohol/drugs abuse. Also age detection should work. So, it can be used within legal and acceptable range.
While I also took the parent comment as sarcasm and see a valid critique in there, what's really ridiculous here is your comment.
Determining age, race and gender from an image is possible in the majority of cases. It becomes problematic when these are used as an indicator for predicting other variables (like IQ). This is a very real problem, and your argumentation just distracts from that.
> It's totally possible to estimate ... race from a face.
It's possible to estimate people into groups of stereotypes as determined by the people who created the models. But race/ethnicity? How can one do that when there is no formal taxonomy? Take the DeepFace example of how they are dong this.[0] They use FairFace to train on. FairFace "labels are East Asian, Southeast Asian, Indian, Black, White, Middle-Eastern and Latino-Hispanic." But the DeepFace guy adds "Merging both east and southeast Asian races into a single Asian race would be better."
This one can stereotype people into "Indian, Black, White, Middle-Eastern and Latino-Hispanic"; apparently they somehow know through images which Latinos are Spanish speaking and/or from counties related to Spain. And let's just merge east and southeast Asian people together into one "Asian"; never mind that many of these Asian countries are homogeneous societies both ethnically and culturally (and as applies to this, drastically different facially) with varying ethnic sub-groups within them.
All that before even getting started on the US. Genetically, many Americans are "multi-race", even the ones getting grouped into White or Black. This is a touchy subject for folks, but will leave this link[1]; it was so common, some states went out of their way to make sure "invisible blackness" was not getting by them.
In the same vein, Native tribes had their own rules for what to do with their "mixed" children. Lucky for them, if these kids were also "white" (though this is not the grouping DeepFace would put them; then they had a shot at getting land or payments when their tribes were...relocated) I mean there is no taxonomy for Natives in DeepFace so I guess they just go with other brown people here? I would not at all say that it is totally possible to estimate race, unless you are only looking for what will amount skin tone, even when that is not what they think they are doing.
Finally, this part of DeepFace's post on race/ethnicity[0]: "Recognizing ethnicity from face photos could contribute a huge contribution to missing children, search investigations, refugee crisis and genealogy research. " How? Me and a couple of my siblings offspring would be in at least 3 different groups using this (and not based on melanin production, we all look very not-related) . The DeepFace people must live in a very homogeneous society to think their logic on this subject is anything more than stereotyping.
Your 20W meat computer does it in the background without you having any control of this process, is that unethical, too?
Besides, people not being the same is a fact of reality which really shouldn't be suppressed from consciousness too much, but that's perhaps a discussion for a different thread.
Part of the core human experience is to estimate these parameters in social settings. It's how we make friends, evaluate social situations, and navigate life. I can't imagine being told we should wear blindfolds. Why would it not be appropriate for computers to do so?
It's not like we'd be using these algorithms to make hiring decisions. We already have a lot of protective legislation on the books.
Ethically viable? That threshold has been crossed a long time ago when any of these related sets of technocratic tyranny enabling abilities were developed.
There are now cameras everywhere and now facial recognition is being integrated into everything, on top of all the other ways of essentially harvesting human rights and dignity for profit and oppressive purposes.
Your mention of ethical viability seems rather late to the situation, a bit of a feel of closing the barn door after the horses have bolted.
What humanity is creating here is more akin to the Matrix than people are able or willing to even acknowledge, even if it is many years out and turns out even worse than the limited imagination of people from the past. Maybe Her and The Terminator are more immediate phases of human development, but the trajectory is clearly there when you simply look at humanity’s tendencies over history and even now. Just as an example, did you ever think you would see something so nakedly and blatantly homicidal and evil as what the Israelis have done and are even still doing in Gaza? Does any of that indicate that humanity is going to do anything responsible and ethical?
The only purpose of appeals to ethics and morality appear to have been to disarm the dominant power by the subordinate power that aspires to become the dominant power in the world, an amoral, unethical, nakedly narcissistic and psychopathic power.
Love this package. has a bunch of functions for most face recognition, detection and feature extraction purposes. love the readme.md docs for familiarizing with the basic features. PLUS it offers different models(backends) for the tasks which lets you try out a bunch of approach for the same task without writing a custom function
The performance is focused on correctness and the APIs work with individual images. The underlying models can be run with batches, but you need to extract the architecture code to run directly. As a result, while I started with DeepFace, I mostly just used the models.
I'm curious whether others here are working with supervised dimensionality reduction for face embeddings, particularly using single-task or multi-task learning approaches.
While clustering tends to perform better after dimensionality reduction, selecting the optimal dimensions depends heavily on your specific use case. This makes it more complex than simply applying PCA or t-SNE.
If you have labelled data, you can try linear discriminant analysis (LDA; also known as canonical discriminant analysis), which maximizes the between-class variance while minimizing the within-class variance to best separate different classes by projecting data onto a new space that maximizes class separability (for whichever classes help your specific use case).
Very nice package. I used it recently for a project where I needed to detect faces in images as tasks with celery [0]. I wonder if there is an equivalent for OCR.
[0]: https://github.com/srugano/facematch
I somehow miss the time when face recognition was the only "controversial" area of ML/AI.
How do you measure positive and negative societal impact of this technology?
I find mobile phone face unlock so useful, giving every citizen the power to use face recognition could be better than a few people, robots that identify someone and give them lifesaving medication are great (but the opposite, robot assassin can also be created). I guess it comes down to good people building good tools. Humans are generally kind and empathetic
Yet also short sighted hairless apes with all the genetic programming that comes with, for better AND for worse.
how accurate is the age detection? i.e. lets say you have someone who looks much younger than their age, would this model be able to detect the persons actual age?
For most machine learning systems, "performance equivalent to an expert human" is the best you can expect - simply because it's learned from training data labelled by expert humans.
So if a person looks much younger than their age in the judgement of an expert human, I wouldn't expect the model to do any better than that.
You should also know a lot of work in this area relies on photos of celebrities scraped from the internet - which is much easier than getting loads of labelled images of normal people in normal situations, which would be a total hassle practically and legally.
Of course that has some benefits - if you know the celebrity's date of birth and the date of the photo, you don't need to rely on human labelling to know the age of the person in the photo! But it has the major disadvantage that if your application doesn't involve professionally made up people with movie star looks in evening wear on red carpets - you might find real world performance falls short of the benchmark claims.
> For most machine learning systems, "performance equivalent to an expert human" is the best you can expect - simply because it's learned from training data labelled by expert humans.
However, this is one of the few cases where presumably the data could be perfect, far exceeding the ability of a expert human.
Some things have no ground truth. For example, masking objects for training vision object detection models. Does the object end at this pixel or that one?
It depends on the quality of the image. It's most accurate for a square head-on photo, like a passport photo, and can be wildly off for occluded or angled photos. It's more accurate for people under 30, for older people it often underestimates age.
i would assume it CAN Be accurate if people 'look their age' but of course you can easily have someone who looks younger or older than their age that skews the overall etimate
Check out the README. They comment on the accuracy.
Quite curious how no one ever talks about “responsible use of facial recognition” or policies to control the use of facial recognition, as it totally pervades and destroys the ability of a person to remain anonymous, at all.
It’s always curious to me how the peasants always seem eager to facilitate the interests of the monarchs to oppress them rather than their own interests to remain free from control by the narcissistic psychopathy of the ruling class prone to tyranny. What do you do as a peasant once you’ve closed the trap you created and led yourself into? I guess maybe more accurately would be to say that it is the aspirational minor nobility that facilitates the creation of the structure that serves the creation of oppressive, top down structures. It’s an odd human characteristic.
Your criticism seems somewhat misplaced. An open-source facial recognition library enables the peasants to wield the same tools that the monarchs already have at their disposal. The cat is out of the bag, and there’s no putting it back.
We can ban the stuff until such time when we truly need it.
Doesn’t help with the climate crisis? No. Does it help with any of the ongoing health threats? No. Does it help hungry people finding food and healthcare or somehow advancing science or any cultural benefits? No.
Does it consume vast amounts of water and electricity to facilitate “bad example of humanity” use cases? Oh yes.
Just feel our priorities are not where they should be.
So we as humanity should stop all actions that are not directly solving the worlds largest problems (at least if there might be slightly negative side effects)? And what’s up with the energy and water consumption argument. I accept this for LLMs maybe, but not for a local python script that runs on consumer hardware. Should we stop playing video games, too?
Respectfully, I think it's more likely that you're just not personally plugged into and/or have your awareness tuned to them but those conversations are definitely happening. There's not a single consensus (at least in the U.S.), but discussion definitely occurs and in many cases has led to concrete action.
[1] https://www.npr.org/2021/05/07/982709480/massachusetts-pione...
[2] https://www.nytimes.com/2019/05/14/us/facial-recognition-ban...
[3] https://www.wired.com/story/face-recognition-banned-but-ever...
There is the EU AI Act which heavily regulates the use of facial recognition, but I feel that we as “the people in tech who makes these things” should be a lot more conservative in creating frameworks, abstractions and generally advancing/facilitating facial rec use.
It’s even more shocking as this library also incorporates a great deal of cultural bias. e.g. gender, emotion are attributes which vary a lot more than what the models allow for.
> Quite curious how no one ever talks about “responsible use of facial recognition” or policies to control the use of facial recognition
That's because that debate has already been lost starting in about 2000 to 2001. 9/11 was really the last nail in that coffin.
Post 9/11 made everything shitty. I feel bad for people who didn't know life before it; now it's the normal.
I would imagine other milestones for new and improved shittiness is the drug war, 1993 crime bill, prohibition, Woodrow Wilson and WWI, etc.
[flagged]
How do you estimate IQ from a face with any accuracy?
Technically there is average IQ by country site, just google. Not that difficult to get faces by country. Put them together. Of course there are regulations and ethic. But in some cases it should work well and is more or less acceptable. Like on Down syndrome or alcohol/drugs abuse. Also age detection should work. So, it can be used within legal and acceptable range.
I assume the parent comment is sarcasm because any library estimating gender, age, and race from an image of a face is just as ridiculous.
While I also took the parent comment as sarcasm and see a valid critique in there, what's really ridiculous here is your comment.
Determining age, race and gender from an image is possible in the majority of cases. It becomes problematic when these are used as an indicator for predicting other variables (like IQ). This is a very real problem, and your argumentation just distracts from that.
> Age model got ± 4.65 MAE; gender model got 97.44% accuracy, 96.29% precision and 95.05% recall
What? It's totally possible to estimate gender, age and race from a face. Why would you think otherwise?
> It's totally possible to estimate ... race from a face.
It's possible to estimate people into groups of stereotypes as determined by the people who created the models. But race/ethnicity? How can one do that when there is no formal taxonomy? Take the DeepFace example of how they are dong this.[0] They use FairFace to train on. FairFace "labels are East Asian, Southeast Asian, Indian, Black, White, Middle-Eastern and Latino-Hispanic." But the DeepFace guy adds "Merging both east and southeast Asian races into a single Asian race would be better."
This one can stereotype people into "Indian, Black, White, Middle-Eastern and Latino-Hispanic"; apparently they somehow know through images which Latinos are Spanish speaking and/or from counties related to Spain. And let's just merge east and southeast Asian people together into one "Asian"; never mind that many of these Asian countries are homogeneous societies both ethnically and culturally (and as applies to this, drastically different facially) with varying ethnic sub-groups within them.
All that before even getting started on the US. Genetically, many Americans are "multi-race", even the ones getting grouped into White or Black. This is a touchy subject for folks, but will leave this link[1]; it was so common, some states went out of their way to make sure "invisible blackness" was not getting by them.
In the same vein, Native tribes had their own rules for what to do with their "mixed" children. Lucky for them, if these kids were also "white" (though this is not the grouping DeepFace would put them; then they had a shot at getting land or payments when their tribes were...relocated) I mean there is no taxonomy for Natives in DeepFace so I guess they just go with other brown people here? I would not at all say that it is totally possible to estimate race, unless you are only looking for what will amount skin tone, even when that is not what they think they are doing.
Finally, this part of DeepFace's post on race/ethnicity[0]: "Recognizing ethnicity from face photos could contribute a huge contribution to missing children, search investigations, refugee crisis and genealogy research. " How? Me and a couple of my siblings offspring would be in at least 3 different groups using this (and not based on melanin production, we all look very not-related) . The DeepFace people must live in a very homogeneous society to think their logic on this subject is anything more than stereotyping.
[0] https://sefiks.com/2019/11/11/race-and-ethnicity-prediction-...
[1] https://en.wikipedia.org/wiki/One-drop_rule
None of what you've said disproves the fact that it is possible to estimate race.
You might not think it's a good idea, and it might be very unreliable, but you can't totally do it.
> None of what you've said disproves the fact that it is possible to estimate race.
My 3rd sentence does exactly that.
"How can one do that when there is no formal taxonomy?"
Estimating race obviously does not require a formal taxonomy of race. Why would you think it does?
It's possible as an estimate, sure, but imho that doesn't make it ethically viable.
Your 20W meat computer does it in the background without you having any control of this process, is that unethical, too?
Besides, people not being the same is a fact of reality which really shouldn't be suppressed from consciousness too much, but that's perhaps a discussion for a different thread.
That’s not the same and hopefully you know that and are just recklessly equating incomparable things.
Maybe, but that's not what they claimed.
> ethically viable
Part of the core human experience is to estimate these parameters in social settings. It's how we make friends, evaluate social situations, and navigate life. I can't imagine being told we should wear blindfolds. Why would it not be appropriate for computers to do so?
It's not like we'd be using these algorithms to make hiring decisions. We already have a lot of protective legislation on the books.
Ethically viable? That threshold has been crossed a long time ago when any of these related sets of technocratic tyranny enabling abilities were developed.
There are now cameras everywhere and now facial recognition is being integrated into everything, on top of all the other ways of essentially harvesting human rights and dignity for profit and oppressive purposes.
Your mention of ethical viability seems rather late to the situation, a bit of a feel of closing the barn door after the horses have bolted.
What humanity is creating here is more akin to the Matrix than people are able or willing to even acknowledge, even if it is many years out and turns out even worse than the limited imagination of people from the past. Maybe Her and The Terminator are more immediate phases of human development, but the trajectory is clearly there when you simply look at humanity’s tendencies over history and even now. Just as an example, did you ever think you would see something so nakedly and blatantly homicidal and evil as what the Israelis have done and are even still doing in Gaza? Does any of that indicate that humanity is going to do anything responsible and ethical?
The only purpose of appeals to ethics and morality appear to have been to disarm the dominant power by the subordinate power that aspires to become the dominant power in the world, an amoral, unethical, nakedly narcissistic and psychopathic power.
[flagged]
This looks very similar to what ChatGPT would write
It probably is but the commentator is the creator of the project.
[flagged]
AI comment
Why do people even do this? What's the point in farming HN karma?
Promoting HN posts can earn a lot of money I'd wager. The audience tends to be wealthy.
Edit: A quick search suggests that one HN upvote costs about 9 cents. I wonder why we don't see even more bots.
Hey good post. I enjoy your blog also.