double-take icon indicating copy to clipboard operation
double-take copied to clipboard

Published 20 hours ago • verified publisher icon jakowenko

[BUG] Car bumper recognized as face. Need a way to teach that something recognized as a face is not one.

Open user897943 opened this issue 2 years ago • 6 comments

Describe the bug Detecting car as face. No way to "teach" the detector to ignore this particular "face".

Version of Double Take 1.6.0

Expected behavior It's saying with 98% confidence that a car is a face... this is clearly wrong. Is there a way to teach detectors to ignore "faces"? Basically a function to say "learn this image is not a face". Screenshots 9ac75ec1-634b-4c02-a76d-0562fe001b20 (2)

Hardware

  • OS: Docker on Debian

  • Browser (if applicable) Not applicable

Additional context There are so many false positives and also detections of faces that aren't the same person but with high confidence. There should be a function (API permitting) to ask the detectors to learn "this image is NOT this person" so that it would improve detection by ignoring that image. On the screenshot above I have literally hundreds of detections all above 90% confidence because it's night time so the noise in the image triggers detection. It's spamming the matches screen.

user897943 avatar Nov 08 '21 22:11 user897943

Can second that. My cat is not a man 👍

Sangoku avatar Nov 12 '21 07:11 Sangoku

The problem with this is not in the double take app, but in the deepstack or compreface applications or whatever you use..

parhedberg avatar Nov 13 '21 14:11 parhedberg

The problem with this is not in the double take app, but in the deepstack or compreface applications or whatever you use..

This is why I said "API permitting". Do you know if there is an API call for "untraining" a specific "thing" against a name on the database? If there is then DoubleTake should offer that for cases like these when you want to give the API a set of images to tell it that it ISN'T what it thinks it is.

user897943 avatar Nov 16 '21 11:11 user897943

Are you using Deepstack by any chance? If so I would recommend giving Compreface a go instead if you are able to, since moving over I haven't had a single silly detection.

wills106 avatar Jan 21 '22 08:01 wills106

I am on Compare face. It still thinks my cat is a person.... No false training done.

Problem is it is a general problem that you cant train negatives. Almost all AI staks support that feature.

Sangoku avatar Jan 21 '22 09:01 Sangoku

Was there ever a resolution to this? I'm not sure what I need to do to teach double take or Compreface that my mailman is not me...

totalitarian avatar Jul 23 '22 17:07 totalitarian