Face and license blurring (GDPR territories)

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
10 messages Options
Reply | Threaded
Open this post in threaded view
|

Face and license blurring (GDPR territories)

Nick Whitelegg-2
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Nick Whitelegg-2

... sorry, this sentence maybe could be misconstrued. "however I now have a collaborator to work on exploring an open source panos platform."

This is very much a joint-effort project between myself and the person I'm collaborating with, I want to make that clear.

Thanks,
Nick



From: Nick Whitelegg
Sent: 06 October 2020 21:41
To: [hidden email] <[hidden email]>
Subject: Face and license blurring (GDPR territories)
 
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Niels Elgaard Larsen
In reply to this post by Nick Whitelegg-2
Nick Whitelegg:

> Hi,
>
> Apologies if this is only tangentially OSM related, but I thought I'd ask here to try
> and get some expert advice.
>
> As you may know, Mapillary has been bought by Facebook and there has been interest in
> developing, or at least starting to develop/actively researching the possibility of,
> some sort of open source alternative. I have been developing OpenTrailView
> (opentrailview.org), however I now have a collaborator to work on exploring an open
> source panos platform.


That sounds great.

> The main question I have relates to the very necessary privacy steps that must be
> taken, in particular face and license plate blurring. I have experimented with
> various libraries using various datasets and models, and have found that the
> understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which
> advertises itself as something specifically aimed at implementing the privacy
> protections needed to comply with the GDPR, seems to be working the best.
>
> It detects faces and license plates in clear view on panoramas, which can then be
> blurred.
>
> My question, then, is what to do about people, or cars, which are further away from
> the camera? In these cases, the algorithm does not necessarily detect the face or
> license plate, but on the other hand in general the faces and license plates are not
> clearly visible, or identifiable, in any case.
> So in summary, the tool blurs clearly visible faces or license plates, but in general
> does not blur those which are not clearly visible.


Hard to judge without examples.
Understand-ai has two examples on github, but that is very clear images.

You will probably have to let users add and remove blurs.
That is what Mapillary do.

> Apologies once again that this is only tangentially related to OSM (OpenTrailView
> uses OSM to connect panos together, so not completely unrelated) but it is very much
> an open geodata issue, so I thought I'd ask to get feedback.
>
> I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.
>
>
> Thanks,
> Nick
>
>
> _______________________________________________
> talk mailing list
> [hidden email]
> https://lists.openstreetmap.org/listinfo/talk
>


--
Niels Elgaard Larsen

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

SimonPoole

Am 07.10.2020 um 01:13 schrieb Niels Elgaard Larsen:
> ...
> You will probably have to let users add and remove blurs.
> That is what Mapillary do.
>
They do not, they stopped providing that facility literally years ago,
and they've gone as far as no longer storing unblurred images even for a
limited time now.

The other important point to understand is that there is no reason to
believe that face and licence plate blurring is sufficient to avoid
trouble in countries with strict data protection regulation as long as
people and vehicles can be identified and behaviour associated with
individuals can be deduced from the images (with other words blurring
would have to be far more complete to be on the safe side). There is
simply no case law, because there have been, afaik, no cases that have
actually gone to court.

tl;dr version you need to make your own risk assessment (and ask your
own counsel).

Simon


_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk

OpenPGP_0x4721711092E282EA.asc (5K) Download Attachment
OpenPGP_signature (505 bytes) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

cquest
In reply to this post by Nick Whitelegg-2
Le 06/10/2020 à 22:41, Nick Whitelegg a écrit :
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


We have tested blurring using image segmentation which allows to blur full parts of pictures like people and cars, not only faces and license plates.


Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/


The code used is on github: https://github.com/tyndare/blur-persons/


We did some tests using TPU to speedup the process.


-- 
Christian Quest - OpenStreetMap France

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Christian Quest - cquest@openstreetmap.fr
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

SimonPoole

Am 07.10.2020 um 10:25 schrieb Christian Quest:

>
> We have tested blurring using image segmentation which allows to blur
> full parts of pictures like people and cars, not only faces and
> license plates.
>
>
> Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/
>
>
> The code used is on github: https://github.com/tyndare/blur-persons/
>
>
> We did some tests using TPU to speedup the process.
>
>
That looks (IMHO) very good and addresses some of the concerns I pointed
out, naturally not perfect, but it is far better than simply blurring
faces and number plates. As a tendency I would remove any colour of the
blurred object too in particular to make cars even less identifiable.

Simon

PS: why handbags are important for OSM is a bit of a mystery though :-)


_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk

OpenPGP_0x4721711092E282EA.asc (5K) Download Attachment
OpenPGP_signature (505 bytes) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Nick Whitelegg-2
In reply to this post by cquest
Hello Christian,

This does indeed look very nice, it's providing much more extensive blurring than what I've tried so far.

Thanks to everyone also for the replies.

Nick

From: Christian Quest <[hidden email]>
Sent: 07 October 2020 09:25
To: [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Le 06/10/2020 à 22:41, Nick Whitelegg a écrit :
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


We have tested blurring using image segmentation which allows to blur full parts of pictures like people and cars, not only faces and license plates.


Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/


The code used is on github: https://github.com/tyndare/blur-persons/


We did some tests using TPU to speedup the process.


-- 
Christian Quest - OpenStreetMap France

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Nick Whitelegg-2
... to follow up on this, it works great on the one pano I've tested so far - I selected this one because it had a 'not-clearly-visible' face and I wanted to see how it would be handled. There was one adult man and two children in this pano, they're all effectively obscured. The previous blurring tools I used blurred all the faces but they didn't blur the child who was partly looking away (with the face not visible)

Christian - thanks once again for this!


Nick



From: Nick Whitelegg <[hidden email]>
Sent: 07 October 2020 17:31
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Hello Christian,

This does indeed look very nice, it's providing much more extensive blurring than what I've tried so far.

Thanks to everyone also for the replies.

Nick

From: Christian Quest <[hidden email]>
Sent: 07 October 2020 09:25
To: [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Le 06/10/2020 à 22:41, Nick Whitelegg a écrit :
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


We have tested blurring using image segmentation which allows to blur full parts of pictures like people and cars, not only faces and license plates.


Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/


The code used is on github: https://github.com/tyndare/blur-persons/


We did some tests using TPU to speedup the process.


-- 
Christian Quest - OpenStreetMap France

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Nick Whitelegg-2

..sorry, the photo ID in that URL is incorrect, should be 9728, not 9928.

Nick



From: Nick Whitelegg <[hidden email]>
Sent: 10 October 2020 21:37
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
... to follow up on this, it works great on the one pano I've tested so far - I selected this one because it had a 'not-clearly-visible' face and I wanted to see how it would be handled. There was one adult man and two children in this pano, they're all effectively obscured. The previous blurring tools I used blurred all the faces but they didn't blur the child who was partly looking away (with the face not visible)

Christian - thanks once again for this!


Nick



From: Nick Whitelegg <[hidden email]>
Sent: 07 October 2020 17:31
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Hello Christian,

This does indeed look very nice, it's providing much more extensive blurring than what I've tried so far.

Thanks to everyone also for the replies.

Nick

From: Christian Quest <[hidden email]>
Sent: 07 October 2020 09:25
To: [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Le 06/10/2020 à 22:41, Nick Whitelegg a écrit :
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


We have tested blurring using image segmentation which allows to blur full parts of pictures like people and cars, not only faces and license plates.


Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/


The code used is on github: https://github.com/tyndare/blur-persons/


We did some tests using TPU to speedup the process.


-- 
Christian Quest - OpenStreetMap France

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk
Reply | Threaded
Open this post in threaded view
|

Re: Face and license blurring (GDPR territories)

Nick Whitelegg-2

Hello everyone,

Having run Christian's blurrer on around 200-300 images now (not all with people and cars) it does seem to be working quite well, it has only failed to detect people on one pano with two children partly looking away from the camera. Incidentally these were close to the edge of the pano. The faces of the children were vaguely visible. I have not allowed access to this.

Would just like to get some input on the acceptability or otherwise of a few examples. I have (temporarily if need be) enabled access to these panos which are what I'd consider edge-cases.

All are available at

where N is a number, detailed below.

People some distance away from the camera. Not clearly visible. Not detected with any of the three pieces of software I've used for blurring, even Christian's:

N = 9731, 9732, 9771

Several people in a cafe/parking area on the top of a mountain. Some people are detected but people looking away/in the distance are not. Note that things are complicated a little with these in that the input image had already gone through a blurring run.

N = 3068, 3076, 3080

People close by:

N = 3096

Anyway, would be great to get some feedback on these 'edge cases', whether they look reasonably 'safe' to release permanently, on the balance of probability.

Thanks,
Nick


From: Nick Whitelegg <[hidden email]>
Sent: 10 October 2020 21:38
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 

..sorry, the photo ID in that URL is incorrect, should be 9728, not 9928.

Nick



From: Nick Whitelegg <[hidden email]>
Sent: 10 October 2020 21:37
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
... to follow up on this, it works great on the one pano I've tested so far - I selected this one because it had a 'not-clearly-visible' face and I wanted to see how it would be handled. There was one adult man and two children in this pano, they're all effectively obscured. The previous blurring tools I used blurred all the faces but they didn't blur the child who was partly looking away (with the face not visible)

Christian - thanks once again for this!


Nick



From: Nick Whitelegg <[hidden email]>
Sent: 07 October 2020 17:31
To: Christian Quest <[hidden email]>; [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Hello Christian,

This does indeed look very nice, it's providing much more extensive blurring than what I've tried so far.

Thanks to everyone also for the replies.

Nick

From: Christian Quest <[hidden email]>
Sent: 07 October 2020 09:25
To: [hidden email] <[hidden email]>
Subject: Re: [OSM-talk] Face and license blurring (GDPR territories)
 
Le 06/10/2020 à 22:41, Nick Whitelegg a écrit :
Hi,

Apologies if this is only tangentially OSM related, but I thought I'd ask here to try and get some expert advice.

As you may know, Mapillary has been bought by Facebook and there has been interest in developing, or at least starting to develop/actively researching the possibility of, some sort of open source alternative. I have been developing OpenTrailView (opentrailview.org), however I now have a collaborator to work on exploring an open source panos platform.

The main question I have relates to the very necessary privacy steps that must be taken, in particular face and license plate blurring. I have experimented with various libraries using various datasets and models, and have found that the understand.ai Anonymizer (https://github.com/understand-ai/anonymizer), which advertises itself as something specifically aimed at implementing the privacy protections needed to comply with the GDPR, seems to be working the best.

It detects faces and license plates in clear view on panoramas, which can then be blurred.

My question, then, is what to do about people, or cars, which are further away from the camera? In these cases, the algorithm does not necessarily detect the face or license plate, but on the other hand in general the faces and license plates are not clearly visible, or identifiable, in any case.

So in summary, the tool blurs clearly visible faces or license plates, but in general does not blur those which are not clearly visible.

Apologies once again that this is only tangentially related to OSM (OpenTrailView uses OSM to connect panos together, so not completely unrelated) but it is very much an open geodata issue, so I thought I'd ask to get feedback.

I am in the UK and the server is in Germany (Hetzner), so GDPR would apply.


Thanks,
Nick


We have tested blurring using image segmentation which allows to blur full parts of pictures like people and cars, not only faces and license plates.


Here is the result: https://takeitout.cquest.org/photo/cquest/blurred/


The code used is on github: https://github.com/tyndare/blur-persons/


We did some tests using TPU to speedup the process.


-- 
Christian Quest - OpenStreetMap France

_______________________________________________
talk mailing list
[hidden email]
https://lists.openstreetmap.org/listinfo/talk