Re-identification

Re-identification answers whether two sightings, perhaps from different cameras or minutes apart, are the same individual object or person. Rather than assigning a fixed class, ReID compares learned appearance representations and decides which observations belong to the same identity. It must cope with changes in pose, lighting, viewpoint, and partial occlusion.

It is what lets a multi-camera system maintain consistent identities across gaps in coverage.

A model maps each detected object or person to an embedding, so that images of the same identity land close together and different identities land far apart, typically learned with contrastive or metric-learning objectives. At query time, a new observation is matched to known identities by embedding similarity. ReID is frequently paired with tracking to reconnect objects after they are occluded or leave and re-enter the frame.

Real deployments span many cameras and long time spans, and ReID is what keeps identities consistent across them, which is essential for multi-camera tracking and analytics. Because it is embedding-based, it connects directly to the representation-learning and similarity-search techniques used across visual AI.

Classification assigns a fixed category, while ReID matches specific instances by appearance similarity, even for identities never seen during training.

Tracking follows objects within a view, and ReID helps re-associate them across cameras or after occlusions, keeping identities consistent.

Multi-object tracking (MOT), Object tracking, Image Embeddings, Embedding and semantic search.