search

minimaxir / person-blocker

Automatically "block" people in images (like Black Mirror) using a pretrained neural network.
Other
2.02k stars 212 forks source link

ethics? ethical uses? #1

Open ghost opened 6 years ago

ghost commented 6 years ago

what are the ethical uses for something like this? is there a way to prevent this from being misused? I imagine it won't be too difficult soon to actually have AR that does this.

Orinion commented 6 years ago

You could use it to make potentially nsfw images sfw! However that would require you to have a trained model for that purpose...

abhi-jha commented 6 years ago

@Orinion Fine. I will do it.

sadovnychyi commented 6 years ago

Privacy. I have a helmet camera which isn't really legal due to privacy reasons, but should be fine if I remove people from it (as google street view does). Curious if raspberry pi will be able to process a single frame in 30ms.

ThoughtfulDev commented 6 years ago

@sadovnychyi Nope a RPi won't. I tried to do it with a video livestream but it wouldn't even be close to 30ms per image. (CPU was a Ryzen 1700x)

ghost commented 6 years ago

@ThoughtfulDev how do the specs compare to a latest model iOS or Android device? I have a RPi 2, not a 3b but I thought the 3b was getting closer to single frame in 30ms?

ThoughtfulDev commented 6 years ago

@rudolfolah I ran the following command and it took ~14.71 seconds on a Intel i5-3320M. python person_blocker.py -i images/img1.jpg -o "person" I just blocked every person in the example Image. I timed the execution right after loading the weights so it took 14 seconds to load the picture,run the prediction, noise the persons and save the picture (not the gif).

So i think a GPU would make this way faster and a RPi would cripple when running this i guess.

Samin100 commented 6 years ago

Running a segmentation net in real-time on a mobile device (such as a helmet) would require something like an Nvidia Jetson TX2 . Even then, I'm unsure if it would be able to perform segmentation at 30 FPS realtime. Maybe 2 TX2s computing in parallel? Maybe 3? I'd love for someone to try it. Seems like it'd be a cool weekend project.