Given that (Anchor, Positive) are images of the same person, and Negative is image of different person, it was said that we should not generate training dataset by collecting random images. This is because, the distance (Anchor, Negative) would be very large and it would easily satisfy the condition

distance(Anchor, Positive) + margin <= distance (Anchor, Negative)

It was suggested that we choose Anchor, Positive and Negative such that distance(Anchor, Positive) is very close to distance (Anchor, Negative) .

Since Negative is obviously the image of different person from the (Anchor and Positive), I wonder how would we achieve the case ‘distance(Anchor, Positive) is very close to distance (Anchor, Negative)’. Do we mean that Positive and Negative should be images of person who are not same but look very similar?

The training set used in the lecture slide also used image of person that looked different for Positive and Negative. In that case the distance (Anchor, Negative) may be larger than distance(Anchor, Positive).