0

I am making a CNN model for face recognition in which I have used Mobilenet. While training I got the val_acc of as high as 90 percent but when I test it on a real video from the webcam or some test video the model performs poorly. What could be the possible reason? I have used the MobileNet without the top layer and have added my own FC layers to it. Then I start training the model.

Mayank Goel
  • 11
  • 2

1 Answers1

0

Try different state-of-the-art face recognition models first. Because MobileNet is a simpler model among those Facenet or VGG-Face.

I recommend you to run VGG-Face or Facenet.

!pip install deepface
from deepface import DeepFace
models = ["VGG-Face", "Facenet", "OpenFace", "DeepFace", "DeepID"]
DeepFace.stream("face_database", model_name = models[0])

If those face recognition models underperform, then your try to process high resolution web cam.

johncasey
  • 1,250
  • 8
  • 14
  • I have also used the VGG19 model but the results were the same. In fact, in this case the val_acc was .9905 but on real video it performed very poorly. Is VGG_Face different? – Mayank Goel Jun 29 '20 at 13:32
  • Model structure is same in both VGG-19 and VGG-Face but VGG-Face is trained with just face photos. That's why, it comes with a huge accuracy in face recognition tasks. – johncasey Jun 30 '20 at 10:14
  • Ohk! Let me try. – Mayank Goel Jun 30 '20 at 13:42