0

I have trained FisherFaces algorithm with about 50 images for two individuals. It does recognize them accurately, but it also recognizes others as them.

I've tried to use the threshold value as an indicator but unfortenately I get also accurate predictions (small distance) on these samples also.

In order to try and battle with this issue I've added some pre processing:

  1. decreased the face area crop on both training and test images
  2. brightened the test images by x scalar (30 was the best value)

I have also considered training a new category with the "unknown" label with random people as the training images.

serj
  • 508
  • 4
  • 20

1 Answers1

1

This is one of the well known problem in the realm of statistics and neural nets called Open Set Recognition Problem.

In essence, the classifier you trained only computes class probability of a sample and generally do not have the ability to discriminate labels outside the known/trained classes.

In the case of face recognition, we need the neural network to have the ability to not only define a decision boundary but also the ability to discriminate data points well.

One of the possible solutions is to give bayesian properties to the neural network. In a bayesian neural network, all weights and biases have a probability distribution attached to them. You can do multiple forward pass during inference time and interpret the output probability of uncertainty of every classes.

Hope my 2 cents helps.

Rex Low
  • 2,069
  • 2
  • 18
  • 47