double-take
double-take copied to clipboard
jakowenko
Question on training and false positives.
I'm using CompreFace with Double Take. Is there some method of marking a detection as being incorrect that helps train the faces?
I started with about 20 photos of myself that I trained it with. Since then I've been using any images of myself that DT generates, and also training them. But I'm wondering if I'm screwing it up by using images that are too small. So I'm wondering if these small images should also be used to train the model. I'm getting people coming to my front door that look nothing like me that are getting detected as me with a confidence of 95%+ that look nothing like me.
Any help is appreciated. Thanks.
may try a different model with a custom compreface docker?
-> https://github.com/jakowenko/double-take/issues/151
Hey @julianrinaldi, thanks for checking this project out. According to the CompreFace devs there is no way to tell the detector an image was a false positive. This was the comment from them a few weeks ago when I asked.
This is technically impossible, we take the face recognition model and use it for all recognitions. You still can retrain this model by yourself and then build CompreFace using your model. Probably we can do something to simplify this process.
I would suggest trying to use the best quality images as possible for the training data, try not to use any from your cameras because the resolution of those probably aren't as good as your phones. I typically use selfies taken from my phone for the training images.
Someone also asked a question about the best way to train the models over on CompreFace's discussion board. It may be helpful to you.
you may try "SubCenter-ArcFace-r100-gpu" docker from compreface. they say thats the heaviest modell and most accurate. on the same computer with the normal compreface model (facenet) my whole system stood still for moments... may you have the powers to test this for your needs...
@jakowenko Thanks for your help. I retrained with just selfies, and the detection is much better. I also changed the unknown confidence level to the same as the main confidence level, so it's marking random people at my door incorrectly as myself much less frequently. Thanks for a great app!
@ozett Thanks for the info. I tried the non-GPU ArcFace on a computer with an i7-9700 and the detection time was 2-3 times as long with no significant increase in the level of accuracy. In all honesty, it seemed to lower the detection accuracy a bit.
The have 1) facenet (facenet-lib), 2) mobilenet(insightface-lib), 3) arcface (insightface-lib).
facenet seems the default and does its job. i had high hopes for more precision with mobilenet, but it detected less faces on my training images than facenet. arcface was really slow and my system stood still and did nothing. also disappointing.
but still more accuracy would be great. more tweaking on insightface directly? or something else?