muliti-similar loss

General Pair Weighting (GPW) frame work

A unified weighting view of sampling problem of metric learning. This indicates that a pair-based method can be formulated as weighting of pair-wise similarities.

Conclude three type of similarities. (Assume we are research on a negative pair)
Similarity-S: Such as the distance between pair.

Similarity-P: Compared with other positive pair

Similarity-N:Compared with other negative pair

multi-similarity (MS) loss

$\alpha$ and $\beta$ and $\lambda$ are hyper-parameters.

It’s partial derivative is

each weight include Similarity-S and N.

there are two steps,
(1)Pair mining
A negative pair is compared to the hardest positive pair (with the lowest similarity), while a positive pair is sampled by comparing to a negative one having the largest similarity
The positive pair is sampled in the same way.
Use Similarity P

(2) Pair weighting