Picking Hard Triplets

For Face Recognition training, Andrew mentions using “hard” triplets in order to enable the NN learn a useful function. Hard triples are examples where d(A,P) is similar to d(A, N). In practice, how do we go about picking hard triplets. One example is using A and N who are very similar looking siblings. But in general, how do we pick “hard” triplets when we have a say 10,000 images of 1000 people?

Hello @Aditya_Ranganath,

“Online triplet loss” / “Online triplet learning” are useful keywords for some readings. This, for example, came first in my google search, but you may need to find more.

Good luck for your search!


1 Like

I believe this is done by human inspection, if the eye has it hard to find the differences then it would be hard for the DL model too, after all its trained by human input.

Visual inspection and choice this is how it should be done.

@rmwkwok Thanks Raymond!

@gent.spah Thanks!

1 Like