"Paper notes" Margin Sample Mining loss:a Deep learning Based Method for person re-identification

Summary

Person Re-identification (ReID) is a important task in computer vision. Recently, deep learning with a metric learning loss have become a common framework for ReID. In this paper, we propose a new metric learning loss with hard sample mining called margin smaple mining loss (MSML) which can achieve better accuracy compared with other metric learning losses, such as triplet loss. In experiments, we proposed methods outperforms most of the State-ofthe-art algorithms on Market1501, MARS, CUHK03 and CU Hk-sysu.

Pedestrian recognition is a very important task in the field of computer vision. The technology of deep learning based on metric learning method has become the mainstream method of Reid nowadays. In this paper, we propose a new metric learning method that introduces difficult sample sampling, which is called MSML. The experiment shows that the method we propose defeats most of the current methods, and obtains state-of-the-arts results on market1501,mars,cuhk03 and CUHK-SYSU datasets. Method

Triplet loss is a very common metric learning method, while quadruplet loss and Trihard loss are two of its improved versions. Quadruplet loss considers the absolute distance between positive and negative samples relative to triplet loss, while Trihard loss introduces the idea of hard sample mining, and MSML absorbs these two advantages.

The goal of metric learning is to learn a function g (x): Rf→rd g (x): \mathbb{r}^f \rightarrow \mathbb{r}^d, which makes the semantic similarity of RF \mathbb{r}^f spatially reflected in RD \mathbb{r}^ The distance from the D space.
Usually we need to define a distance metric function d (x, y): Rdxrd→r d (x, y): \mathbb{r}^d \times \mathbb{r}^d \rightarrow \mathbb{r} to represent the embedded space (embedding spaces ), and this distance is also used to re-identify pedestrian images.

The three-tuple loss, four-tuple loss and trihard loss, which are introduced in domestic and foreign research, are typical metric learning methods. Given a ternary group {a,p,n} \{a,p,n\}, the ternary loss is expressed as:
lt= (da,p−da,n+α) + l_t = (d_{a,p}-d_{a,n}+\alpha) _+
The loss of ternary group only takes into account the relative distance between positive and negative sample pairs. In order to introduce the absolute distance between the positive and negative pairs, a four-tuple loss is added to a negative sample consisting of four-tuple {a,p,n1,n2} \{a,p,n_1,n_2\}, while the four-tuple loss is also defined as:
Lq= (Da,p−da,n1+α) + + (da,p−dn1,n2+β) + l_q = (d_{a,p}-d_{a,n1}+\alpha) _+ + (D_{a,p}-d_{n1,n2}+\beta) _+
If we neglect the effects of parametric α\alpha and Β\beta, we can represent a more general form of four-tuple loss:
L