Security cameras flip to infrared at night, and matching the same person between a color shot and a night-vision one is hard because the color cues vanish. Doing it with no labels telling you who is who is harder still. The common tactic is to slowly guess identities and pair them up, but two things go wrong: the system leans on a vague whole-body impression, and once it makes a bad guess, that error feeds forward unchecked.

SSRL turns that one-way pipeline into a loop that catches its own mistakes. It breaks a person into body-part pieces that act as stable landmarks instead of trusting the fuzzy silhouette, then rebuilds shared identity prototypes each training round and feeds them back to filter out bad labels before the next clustering step. On standard benchmarks it holds its own against the best methods, and on one it beats some approaches that had labels to work with.

That is based on the abstract, so consult the paper for the architecture and full results.