3 ms·

Show HN: Cover Faces from Protest Photos

11 points by aliabd 6y ago

noeltock 6y ago
Well done! Feels like it's missing the larger opportunity to protect other protests globally (by using a more generic/pixelated cover).
- aliabd 6y ago
  Thanks! Absolutely, maybe I'll add the ability to choose or upload an emoji/cover?
rjeli 6y ago
nice. I made something similar a few weeks back at http://deafmute.me http://deafmute.me. It’s a web app that accesses your camera and covers up faces with tfjs, and you can record video by screen recording. tfjs is pretty demanding though, so it’s pretty low FPS on older phones, and I never got around to writing a Kalman filter so that flickers wouldnt expose faces for a few frames.
- aliabd 6y ago
  Just checked it out! Looks good :) Yeah, tfjs is pretty heavy on the browser, but at least you don't need to process on a server. Try using https://gradio.app https://gradio.app if you want to keep processing on a server/your own hardware.
aliabd 6y ago
I used Gradio[1] to create and host this interface. Original model[2] developed by Stanford ML Group and LSC-CNN[3]. [1]: https://github.com/gradio-app/gradio https://github.com/gradio-app/gradio [2]: https://github.com/stanfordmlgroup/blm https://github.com/stanfordmlgroup/blm [3]: https://github.com/vlad3996/lsc-cnn https://github.com/vlad3996/lsc-cnn
Yetanfou 6y ago
It would be even better if the tool could make a difference between legitimate protesters and those who are there to riot, loot, pillage and burn since I see no reason to mask their faces.
- verdverm 6y ago
  Interesting but difficult. What would one do about the rioters wearing "press" on their clothes? Seems double edged considering the murder / suicide we saw this week in Kenosha
  - Yetanfou 6y ago
    For human observers it is rather easy to separate the rioters from protesters when observing video streams. On still images this is much harder unless the image happens to catch a rioter in the act of slinging a Molotov cocktail or swinging a 2x4 at a shopkeeper. A combination of facial recognition and a model which has been trained to recognise rioting and rioters might get you a long way. ...like they're doing in China, except that they're looking for dissenters instead of rioters... ...which is not exactly an endorsement.
    - burfog 6y ago
      It's not easy. The protester seen from the front in one video could also be a murderer seen from the back in a different video. Harmless protesters are safe in the USA. They don't need to be covered up. Covering them is effectively an acknowledgement that they are not in fact peaceful.
      - Yetanfou 6y ago
        Well, yes, in the USA this is probably the case. The world is a lot bigger than the USA and encompasses a multitude of countries where protesting is best done anonymously. The 'black block' tactics by the likes of AFA/Antifa are another story, these are people who go out with the intention to harm those who they consider to be their opponents, i.e. the opposite of the term 'peaceful protests' which is bandied around so much. They are the ones I'd like to unmask or at least not have masked by any software made to protect the privacy of real protesters. On the subject of that protester from one video being a murderer on another, not really. If you only have a single source this problem exists and is unavoidable, so much is clear. When you have multiple sources from different angles it becomes possible to correlate the person seen from the back in source A with the murderer in source B and with that identify that person in source A as the murderer.

11 points by aliabd 6y ago

Well done! Feels like it's missing the larger opportunity to protect other protests globally (by using a more generic/pixelated cover).

aliabd 6y ago
Thanks! Absolutely, maybe I'll add the ability to choose or upload an emoji/cover?

rjeli 6y ago

nice. I made something similar a few weeks back at http://deafmute.me http://deafmute.me. It’s a web app that accesses your camera and covers up faces with tfjs, and you can record video by screen recording. tfjs is pretty demanding though, so it’s pretty low FPS on older phones, and I never got around to writing a Kalman filter so that flickers wouldnt expose faces for a few frames.

aliabd 6y ago
Just checked it out! Looks good :) Yeah, tfjs is pretty heavy on the browser, but at least you don't need to process on a server. Try using https://gradio.app https://gradio.app if you want to keep processing on a server/your own hardware.

aliabd 6y ago

I used Gradio[1] to create and host this interface. Original model[2] developed by Stanford ML Group and LSC-CNN[3]. [1]: https://github.com/gradio-app/gradio https://github.com/gradio-app/gradio [2]: https://github.com/stanfordmlgroup/blm https://github.com/stanfordmlgroup/blm [3]: https://github.com/vlad3996/lsc-cnn https://github.com/vlad3996/lsc-cnn

Yetanfou 6y ago

It would be even better if the tool could make a difference between legitimate protesters and those who are there to riot, loot, pillage and burn since I see no reason to mask their faces.

verdverm 6y ago
Interesting but difficult. What would one do about the rioters wearing "press" on their clothes? Seems double edged considering the murder / suicide we saw this week in Kenosha
- Yetanfou 6y ago
  For human observers it is rather easy to separate the rioters from protesters when observing video streams. On still images this is much harder unless the image happens to catch a rioter in the act of slinging a Molotov cocktail or swinging a 2x4 at a shopkeeper. A combination of facial recognition and a model which has been trained to recognise rioting and rioters might get you a long way. ...like they're doing in China, except that they're looking for dissenters instead of rioters... ...which is not exactly an endorsement.
  - burfog 6y ago
    It's not easy. The protester seen from the front in one video could also be a murderer seen from the back in a different video. Harmless protesters are safe in the USA. They don't need to be covered up. Covering them is effectively an acknowledgement that they are not in fact peaceful.
    - Yetanfou 6y ago
      Well, yes, in the USA this is probably the case. The world is a lot bigger than the USA and encompasses a multitude of countries where protesting is best done anonymously. The 'black block' tactics by the likes of AFA/Antifa are another story, these are people who go out with the intention to harm those who they consider to be their opponents, i.e. the opposite of the term 'peaceful protests' which is bandied around so much. They are the ones I'd like to unmask or at least not have masked by any software made to protect the privacy of real protesters. On the subject of that protester from one video being a murderer on another, not really. If you only have a single source this problem exists and is unavoidable, so much is clear. When you have multiple sources from different angles it becomes possible to correlate the person seen from the back in source A with the murderer in source B and with that identify that person in source A as the murderer.