Certainty about howdy HOT 29 CLOSED

Basically the program went from working almost 100% of the time to <10%.

from howdy.

I also tried removing and remaking models and using just 1 model.

from howdy.

Why did I update lol. Like the number 1 rule if it isn't broke don't fix it.

from howdy.

I've added 2 new important config values. end_report gives a detailed report after a succeeded scan, including the used resolution and winning certainty. max_height scales down the frames from your webcam. This gives a dramatically higher fps, but can make scanning harder.

You can find the config in /lib/security/howdy/config.ini. Try setting it to a higher max and upping the certainty to 6.

from howdy.

changing the max_height does not seem to anything as the output of adding a face is the same Adjusting resolution from 384x288 to 340x374.

from howdy.

In fact changing any value located in the config /lib/security/howdy/config.ini seems to have no effect on the program.

from howdy.

That's not where the config makes the change though, the add command will always capture in full resolution as it makes the models more reliable. You could try renaming the config.ini to something else and look if the program still runs. I'd recommend not adding more than 3 models as every additional model slows down the script a lot.

I think your models are just not accurate enough in some way. The actual capture and compare code hasn't changed a lot between these 2 releases.

from howdy.

I didn't think it had changed. I tried adding one because before I had 2. I just don't get why it is now not wanting to work.
The biggest problem is trying to add a model it just doesn't detect a face to capture

from howdy.

A retry prompt would be really useful here as it is a pain to keep typing in the name of the file every time it fails.

from howdy.

If it's having so much trouble finding you in the add command it will have even greater issues when trying to authenticate you. I'll look into a test function though, so you can see more clearly what's going wrong. It could be that you're just not illuminated enough.

from howdy.

But it was working perfectly and I haven't changed anything other than update Ubuntu and download the newest version of howdy

from howdy.

Like if it just never worked then ok sure maybe it's just buggy on my hardware or something. Though I used it flawlessly for about a day or two. It was a bit slow at first but that may have been my imagination cause it was definitely quicker and more convenient than typing my password everytime after that. It also had a higher certainty value of 3

from howdy.

I am gonna uninstall everything and try again. I think honestly it is the pip programs being screwy again. Also if the new version of howdy installs them automatically it is probably best i uninstall anyway.
Side Note: sorry for so many messages I will try to condense my thoughts or edit a message

from howdy.

I tried uninstalling all the libraries manually it didn't help.

from howdy.

I think the add command (the former learn.py) needs a serious rework. In the meantime i've pushed a new test command to the dev branch. Could you try it and see if it can detect your face?

To download the lastest dev commit:

sudo -i
cd /lib/security/howdy/
git fetch
git checkout dev
pull

You should now have a test command which you can run anywhere with sudo howdy test.

from howdy.

Whelp.
`Opening a window with a test feed

Press ctrl+C in this terminal to quit
Click on the image to enable or disable slow mode

No protocol specified
Unable to init server: Could not connect: Connection refused

(Howdy Test:2083): Gtk-WARNING **: cannot open display: :0
`

from howdy.

Nvm that is an issue with ubuntu using wayland on 17.10. Yes it detects my face. At least I am assuming that the red circle around my face means. I still cannot add myself for detection though.

from howdy.

If the test command can see you it must be a timing issue with your emitters. I've pushed yet another update to dev with a revamped add command, try pulling it and see if it works better.

from howdy.

Congratulations I can scan my face. However authentication is super slow. Like upwards of almost 10 seconds sometimes before the ir camera even starts flashing.

from howdy.

Nvm the slowness was my cpu governor being set to 30 percent. I forget how much horsepower that this deep learning stuff takes.
Looks like me being able to scan my face is working though

from howdy.

Nope I spoke too soon
sudo -i Face detection timeout reached

from howdy.

I can't really do much about that, if the test command shows it's seeing a face then it's just not recognizing you. You could try upping the certainty and adding more models. I'd recommend moving your current models folder to a safe location (just in case).

from howdy.

Well where would you move them? Also the program is mostly working now. Sometimes it fails but unfortunately I have not had much time to play around with it. I am gonna close this issue though. While the program is definitely buggy it is working.

from howdy.

Anywhere is fine, it's just a backup. This script could possibly be much more reliable with an algorithm that's specifically made for black and white IR pictures. It would take ages to train that tough and it has so few users that it's just not worth it at the moment.

from howdy.

Well I would be more than willing to provide my hardware for testing and training this if it get's me closer to a better login or go even further and combine both ir and color images for better detection. However it would be really nice if the I could get more feedback from failed attempts. Such as the score I got compared to what I need to achieve successful authentication. It may also be helpful to save those images for later tweaking of the algorithm. I mean with everyone and their brother adding a connect into everything from phones to laptops. The actual user base may limited at the moment. However the technology is very powerful. Like imagine this linked with something such as OpenPGP smartcards for two factor authentication. Now this is assuming Howdy works properly as intended. Though I still have fun playing around with it.

from howdy.

Maybe one day. As matching IR imagery and the visible spectrum images together is probably one of the hardest things for AI and DNN right now.

from howdy.

The issue is that we would need a lot of training images. Normal photos in the visible spectrum are easy to find, but IR photos are really different.

from howdy.

Yes but finding people for training images becomes a lot easier once you have tools available to do so.

from howdy.

What do you propose as tools?

from howdy.