Training to detect closed eyes

[UlrichVonRekkenin]

Hi there,

1. Is it a good idea to use this one example or modified version of another one to train net to detect closed eyes?
2. How to get more then 8*8 fhog image
   
3. Because train_object_detector.cpp from the same training.xml (produced by imglab) generate something like this one 13*11
   
4. I tried to use frontal close-up images with eyes (strictly horizontal)
5. in the imglab mark up left and right eye
6. used flip

When I dlib::deserialize and process image with opened eye, them detected. But the closed eyes detected too.

[davisking]

HOG isn't going to be super great at telling the difference between closed eyes vs open eyes. I would use a CNN to do that kind of thing if you want to get something accurate.

[UlrichVonRekkenin]

Thanks for your reply!

Something like this CNN, with the final fc<1> to get the probability and loss_binary_log?

[arrufat]

To train a neural network to detect closed eyes, I'd start with this
http://dlib.net/dnn_mmod_ex.cpp.html

Just make sure to read all comments thoroughly, and label your data consistently.