Existing person re-identification (re-id) methods mostly exploit a large set
of cross-camera identity labelled training data. This requires a tedious data
collection and annotation process, leading to poor scalability in practical
re-id applications. On the other hand unsupervised re-i