Read the statement by Michael Teeuw here.
Face recognition recognizing anyone as me
-
@saifkazi How many training images did you use? And did you use training images taken from the actual camera or a bunch of existing photos?
I got decent results when I trained the engine with about 80 images, all taken from the camera. I took the training images in similar lighting conditions to where my mirror ultimately ended up being mounted, and at various distances and angles from the mirror that would be fairly “typical” of someone standing in front of the mirror. I don’t recall the confidence numbers I got, but I had it reliably recognizing myself and my wife, as well as NOT recognizing anyone else and returning to the default profile.
-
@j.e.f.f
Hey Jeff thanks for the reply . I used existing pic taken from my old usb webcam and It’s only one pic. Do you think taking pic from dslr would make any difference? Or should I use existing usb webcam ? -
@saifkazi Use the camera you are using for facial recognition (if that’s your USB cam, then use that), ideally in the same or similar lighting conditions as where your mirror will ultimately be mounted.
Pictures taken with the DSLR will be vastly optically different from what the USB cam will pick up, so you’ll likely have difficulty getting good results.
-
@j.e.f.f
thanks for the reply Jeff .But Some time using fswebcam command for more than 80 image can be frustrating . Still I will try this
Regards Saif -
@saifkazi Indeed, it is quite time consuming. Try taking, say, about 20 pictures to start with to see if your results improve. If they don’t, take another 20 or so. The first 20 will still be there to use for training when you run it the second time.
-
@j.e.f.f
Thanks Jeff for your help
One last thing is can I use other software other than fswebcam ? Like camera on Windows 10 ?because pic quality is quite good in that app it’s good rather than rpi3 fswebcam.Regards Saif
-
@saifkazi I’m not aware of anything else, but I’m not sure the quality is a product of the software so much as how the image is processed in preparation for facial recognition. From what I recall, the image is first lowered in resolution, then converted to greyscale, which would produce a visually low quality image.
-
I lightly modified the training software so that it loops on its own when taking pictures, so that by just standing a few minutes in front of the mirror it gets a lot of different picture.
That might be a solution for you maybe. -
@da4throux
What modifications you made on training software? -
@saifkazi I just checked to see if I still have the code but it’s gone with a module update :(
basically, I think I just modified capture.py in MMM-Facial-Recognition-Tools, and added a loop around the capture.capture(preview) with a wait time of 3 seconds (you get to see if the process was successful in the shell).Hope that help
