I'm quite new in the CNN world.
I'm satisfied with the face-detection-adas-0001 net's performance.
I now wonder what is the best way to detect if the detected faces are mostly frontal or of profile type.
The nets dedicated to the head pose seem to be unable to handle faces that are too much out-of-frontal. Is there a hope in checking the face-detection-adas-0001 layer activation, looking for some "profile face specialized locations" ?
If so, is there a way to convert back the .bin and .xml files to a framework (e.g. tensorflow, keras, caffe?) more suitable for "fine tuning" operations ?
Should I better go towards a frontal/profile classifier?
Best Regards ,
Despite having the prototxt file face-detection-adas-0001 there is no caffemodel file for this particular net (that is very efficient on 16/9 images...) So I can't figure out to go back towards a well suited training framework...
Dear Bigorgne, Erwan,
Also look at https://github.com/opencv/training_toolbox_caffe . Face Detection is one of the models covered. Unfortunately there is no way to "go backwards" from a generated IR file, back to the original model.