Open justadudewhohacks opened 6 years ago
Hi Vincent,
Curious how this can be achieved ?
I thought that for the dlib's face recognizer model it is mandatory to provide the image size to be 150x150.
See https://github.com/davisking/dlib/blob/master/examples/dnn_face_recognition_ex.cpp#L44
Regards & thanks Kapil
Hey kapil, I think the model has just been trained on 150x150 images but you can still pass on images of smaller dimensions. I also used 105x105 images sometimes in face-recogition.js. It doesnt perform as good as 150x150 but still it is possible.
Ok. I had made the assumption that we can not even pass an image smaller than 150x150. Thanks for the clarification.
For some reason the dlib net crashed for me when using images smaller than 105x105, so that was the smallest I was able to use.
Hi Vincent,
I'm wondering, will you implement MobileNetV2 to retrain the face recognition models?
Best,
I was thinking about training a tiny face recognition model at some point, but not sure if I get to do that anytime soon.
Why do you think mobilenet v2 would be a good fit for that?
The MobileNetv2's results lie on the curve more than MobileNetv1 one https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
Hmm, but the current face recognition net is not a mobilenet, it is a simple resnet. Or did you mean the SSD MobileNet v1 face detector?
You are right on the SSD, my bad.
hi Vincent Which size gives the best performance?
Allow smaller sizes of face image inputs then 150 x 150.