Person re-identification

Person re-identification (re-id) is essentially the task of recognizing if a person has been observed in different locations over a set of non-overlapping cameras views. It's an important task for surveillance applications, either for on-line tracking of an individual over a network of cameras, or for off-line retrieval of all videos containing a person of interest.

A typical application scenario of person re-identification: a network of video surveillance cameras monitoring a large public space. The girl is seen at first by the camera in the upper-left corner, then by a second camera of the network (lower-right corner). A person re-identification system should associate these views to the same identity.

[2016,CVPR] Top-push Video-based Person Re-identification

$$ f(D) = (1-\alpha) \sum_{x_i,x_j,y_i=y_j} D(\vec{x_i},\vec{x_j}) +\\ \alpha \sum_{x_i,x_j,y_i=y_j} \max{D(\vec{x_i},\vec{x_j})-\min_{y_k\neq y_i}D(\vec{x_i},\vec{x_k})+\rho,0} $$


Deep Metric Learning for Person Re-Identification


  • DML can learn a similarity metric from image pixels directly. All layers in DML are optimized by the same objective function, which are more effective than the hand-crafted features in traditional methods.
  • The multi-channel filters learned in DML can capture the color and texture information simultaneously, which are more reasonable than the simple fusion strategies in traditional methods,e.g.,feature concate-nation and sum rule.
  • The structure of DML is flexible that can easily switch between view specific and general person re- identification tasks by whether sharing the parameters of sub-networks.

A brief intro of CNN

In machine learning, a convolutional neural network (CNN, or ConvNet) is a type of feed-forward artificial neural network in which the connectivity pattern between its neurons is inspired by the organization of the animal visual cortex, whose individual neurons are arranged in such a way that they respond to overlapping regions tiling the visual field.

CNN works well in "sample label" training data sets. But for person re-identification, the subjects in the training set are generally different from those in the test set. In this work, researcher introduced a siamese CNN.

connection function is used to evaluate the connection between sample pairs, candidates could be Eu- clidian distance, Cosine similarity, absolute difference, vector concatenate and so on.

$$ \begin{aligned} S_{euc}(x,y) &= -\sum_{i}(x_i-y_i)^2\\ S_{cos}(x,y) &= \frac{\sum_i x_i y_i}{\sqrt{\sum_i x_i x_i \sum_i y_i y_i}}\\ S_{abs}(x,y) &= -\sum_{i}|x_i-y_i|\\ S_{con}(x,y) &= \sum_{i} \omega_i [x;y]_i \end{aligned} $$