I was having one doubt when we are using face recognition and getting the probability of being that person or not. Then If we apply the same technique to let’s say a million people then we shall be requiring minimum of a billion to 10 billion images for the training, Even if we get the image how will we process that as all the images need not to be in same shape (dimension).

Other thing is If we will get a million probability as a response then the probability will be close to each other how shall it be segregated ?

1 Like

You can also use pretrained models with transfer learning, so you can obtain already learned information from images of other sorts and fine tune to your model with a limited dataset.

The probabilities depend always on the models accuracy, the probabilities will not be distributed equally and the model chooses the highest one, but could be wrong as its a probability after all.

Hello @Ayush.Agr,

@gent.spah has made a great point about transfer learning - it saves us from training from scratch with many photos.

Besides, from your message, you have correctly pointed out that getting a million probabilities might not be a good idea. In fact, in practice, we do not do it this way - instead of computing 1000 probabilities for 1000 persons, we compute similarity scores between two persons.

The core idea is, you first prepare (A) a representation computed by the model for each person, then as new photo comes in, you compute (B) the representation for that photo, and see which one of (A) is most similar to (B).

For more ideas, I recommend DLS C4 W4. You may audit the course to just watch the videos.. There are 5 videos for it:

If you have not done the rest of the DLS, then don’t worry about any technical specifics that are not familiar to you, and just skim the videos for some ideas. In particular, in “One Shot Learning” video, you will see that it is possible for us to just need one photo per new person.

Cheers,

Raymond

5 Likes

Great reference @rmwkwok, its been a while I done DLS, I forgot about it. Thank you!

1 Like

Thanks a lot I will surely go through it once

1 Like