Re-identification in multi-object tracking

WHAT THIS TESTS Whether you understand that across long occlusions motion models fail and identity must be recovered from appearance, the Re-ID problem.

A GOOD ANSWER COVERS In multi-object tracking, frame-to-frame association by IoU or a Kalman prediction works only while objects stay visible and move predictably. When an object is occluded for many frames or exits and re-enters, its predicted position is stale, so position-based matching breaks. Re-identification is the task of correctly reassigning a reappearing detection to its original track id. The high-level strategy is appearance-based. Train a Re-ID embedding network, often with a triplet or contrastive loss, so crops of the same identity map to nearby vectors and different identities map far apart. For each active track, maintain a gallery of recent appearance embeddings. When a detection cannot be matched by motion, compare its embedding against the embeddings of recently lost tracks using cosine or Euclidean distance; a close match revives that track's id rather than creating a new one. DeepSORT combines this appearance distance with Mahalanobis motion distance.

COMMON WRONG ANSWERS Relying solely on IoU or Kalman prediction, which cannot bridge long gaps. Always spawning a new id on reappearance, inflating identity switches. Comparing raw pixels instead of a learned embedding, which is brittle to pose and lighting.

LIKELY FOLLOW-UPS How is the Re-ID network trained and evaluated? How long should lost tracks be retained before deletion? How does DeepSORT fuse appearance and motion? How do you handle similar-looking objects like uniformed players?

ONE CONCRETE EXAMPLE A pedestrian walks behind a bus for two seconds. The Kalman prediction drifts and the track is marked lost but kept. When the person reappears, a new detection's embedding is compared to the lost track's stored embeddings, the cosine distance is small, so the original id is restored instead of assigning a fresh one.