Glossary / Computer vision
What is re-identification?
Re-identification is matching the same person or object across different cameras, or after they leave and re-enter view, using an appearance descriptor rather than a name. It answers whether two sightings are the same individual without determining who that individual is. This distinction has real consequences for both accuracy and privacy obligations.
A single camera tracking a person across a frame is doing tracking. When that person walks behind a pillar and reappears, or walks out of one camera's view and into another's, tracking alone cannot connect the two. Re-identification does.
It works by converting a person's appearance into a numeric vector, an embedding, that captures clothing colour, texture, body proportions and gait cues. Two sightings are compared by measuring the distance between their embeddings, usually cosine similarity. Above a threshold, the system calls them the same person.
Nothing in that process involves identity. The system holds a vector, not a name.
Re-identification is not face recognition
These get conflated constantly, and the conflation causes both bad procurement and bad compliance decisions.
| Capability | Question answered | Reference data held |
|---|---|---|
| Tracking | Where did this object move within one camera view? | None. Session only. |
| Re-identification | Are these two sightings the same individual? | Appearance embeddings, typically short lived. |
| Face recognition | Which enrolled person is this? | An enrolled gallery linking faces to identities. |
The practical consequence: re-identification needs no enrolment. There is no gallery of known people, no consent flow for enrolment, no identity database. For retail journey analytics, queue measurement and dwell time across zones, that is exactly what you want, and reaching for face recognition to solve those problems imports obligations the task never required.
Where re-identification breaks
Its dependence on appearance is also its weakness, and the failure modes are specific.
- Similar clothing. A school, a factory with uniforms or a hospital defeats appearance matching almost completely. Uniformed environments are the hardest case, not the easiest.
- Wardrobe change. Removing a jacket produces a new identity as far as the system is concerned.
- Time horizon. Reliability falls sharply over hours. Same day matching within a site is a reasonable ask. Matching a visitor today to the same visitor next week is not.
- Lighting and camera differences. The same red shirt under sodium lighting on one camera and daylight on another produces different embeddings.
- Viewpoint. An overhead camera and an eye level camera see very different things.
This is why re-identification quality is reported as Rank-1 accuracy, the share of queries where the correct match is the top result, and mAP, which accounts for all correct matches in the ranked list. Public benchmark figures above 90 percent Rank-1 are common and translate poorly to a real site, because the benchmark does not have your uniforms, your lighting or your camera angles.
The threshold decision
Why the threshold is a business decision, not a technical one
Matching happens when similarity exceeds a threshold. Raise it and you get fewer wrong matches but more journeys broken into fragments. Lower it and journeys stay whole but distinct people get merged.
For footfall and journey analytics, a broken journey undercounts a visit while a merged one invents a path that nobody walked. Fragmentation is usually the safer error, because it biases toward undercounting rather than fabricating. For security use, the calculus reverses. The threshold should be set from the consequence of each error type at your site, and any vendor who ships a single default without asking has not thought about it.
Re-identification and DPDP
It is tempting to conclude that because re-identification does not know who anyone is, it falls outside data protection obligations. That conclusion is unsafe. An embedding that reliably distinguishes one individual from others, and that persists, may constitute personal data even without a name attached, because the individual is singled out. Whether it does depends on how long it is retained, whether it can be combined with other records you hold, and what it is used for.
The defensible position is to treat embeddings as personal data and design accordingly: keep them on site, retain them for the shortest period the analytics genuinely requires, usually hours rather than days, and avoid joining them to any record that carries identity unless you have a lawful basis and have said so in your notice. Where the analytics is purely statistical, discard the embedding once the visit is closed and keep only the aggregate.
Processing at the edge helps here, because the embeddings never leave the premises and the retention window is enforceable in one place. Rule 4 commences on 13 November 2026. This is general information rather than legal advice, and deployments involving employee or resident data should be reviewed by your own counsel.
Frequently asked questions
Does re-identification identify who someone is?
No. It determines that two sightings are the same individual by comparing appearance descriptors. There is no name, no enrolled gallery and no link to an identity record unless one is deliberately added. That is the defining difference from face recognition, and it is why re-identification suits analytics tasks that do not need identity.
Is re-identification biometric data under DPDP?
The safest working assumption is that a persistent appearance embedding capable of singling out an individual should be treated as personal data, even without a name. Whether it is depends on retention, what else you hold and how it is used. Treating it as personal data and minimising retention is the defensible position. This is general information, not legal advice.
How long does re-identification stay accurate?
Within a site on the same day, it works reasonably well. Reliability drops sharply across hours as lighting changes and people remove or add layers of clothing, and across days it is not dependable at all, because the appearance it depends on has changed. Design for same session matching.
Why does re-identification fail in uniformed environments?
Because it matches on appearance, and uniforms deliberately make appearance identical. Schools, hospitals and factories with issued workwear are the hardest environments for it. Where identity matters in those settings, an access control credential or a badge is a more reliable signal than appearance matching.
What are Rank-1 and mAP?
Rank-1 accuracy is the share of queries where the correct match appears as the top result. mAP, or mean average precision, accounts for where all correct matches fall in the ranked list, so it reflects performance more completely. Benchmark figures for both translate poorly to a specific site with its own lighting and camera angles.
Can I use re-identification instead of face recognition for retail analytics?
For journey, dwell and queue analytics, usually yes, and it is the better choice. Those questions are about paths and durations, not identities, so re-identification answers them without enrolment, without an identity gallery and without the obligations that come with holding facial data.
Need cross camera analytics without facial data?
IndoAI runs re-identification on site, so appearance descriptors never leave your premises. Useful where you need journey or dwell analytics but do not want to hold biometric data.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile