Person reidentification:

* Given: Network of non-overlapping cameras
* Goal: identify when a new individual seen in camera A is the same as a
previously seen individual in camera B. 

In re-id terminology there is a gallary of known images and there is a set of
probe cameras.

Cameras are low resolution
There are viewpoint, illumination, occlusion, background clustter, blahblah normal stuff

There are inter-camera variations. 

Usually a distance metric M is learned to find a distance between 
two vector representations of a person. 
This distance metric is typically some mahalanobis distance matrix.

A track of images from a person establishes multiple pictures of the person.
Can compute a single feature vector from multiple images to represent a single
person.