Foundations7 min read
Open-Set Person ReID: Rejection, Unknowns and Evaluation
Learn how open-set Person ReID retrieves candidates, rejects unsupported matches, handles unknown identities and evaluates false accepts with human review.
Your gallery contains 500 people. The next query belongs to person 501.
The search engine still returns a first result—it has to. Perhaps the candidate wears the same dark suit and earns a score that looks reassuring on screen. But nearest is a position in a list, not a verdict.
That is the central problem of open-set Person ReID. A useful system needs two abilities: retrieve plausible candidates and reject the entire gallery when none is sufficiently supported. It may return “no supported match,” ask for review or create a new identity scoped to that project. If you automatically accept top-1, every unknown person becomes an opportunity for a false match.
The Person ReID guide explains the complete crop-to-retrieval pipeline. This article covers the open-set decision structure; score calibration itself belongs in the Person ReID similarity threshold guide.
Remove one assumption and the task changes
Standard ReID training often already uses identities that do not appear in the test set. The model learns an embedding intended to generalize to new people. Yet many standard evaluation protocols remain closed-set at query time: every valid query is assumed to have at least one matching identity in the gallery.
Open-set evaluation removes that guarantee.
| Setting | Is the query identity guaranteed to be in the gallery? | Required output |
|---|---|---|
| Closed-set retrieval | Yes | Rank the gallery; a correct candidate exists |
| Open-set retrieval | No | Rank candidates and decide whether any match is supported |
| Open-world operation | No, and the gallery can evolve | Reject, review, enroll, update and delete over time |
The early Open-set Person Re-identification benchmark described the problem as two connected tasks: detection of whether the probe person is present in the gallery, then identification of the corresponding gallery person when accepted. That remains a useful mental model.
“Unknown” is a relationship to this gallery
An “unknown” outcome means not represented in this gallery under this policy. It does not mean that all unknown people form one coherent class.
Imagine a project gallery containing Alice, Ben and Chen. Dina, Elias and Fatima are all unknown to that gallery, but you should not push them into one shared embedding cluster called unknown. They are different people whose only common property is absence from the current enrollment set.
This distinction prevents three design errors:
- training one catch-all unknown class from an arbitrary sample;
- interpreting rejection as proof that two rejected queries are the same person;
- treating “unknown here” as “unknown everywhere.”
Gallery scope matters. A participant can be known within one authorized event and unknown in another project whose gallery is deliberately isolated.
Keep the ranking separate from the decision
A practical pipeline can keep five stages visible:
one-person query crop
→ versioned appearance embedding
→ nearest-neighbor retrieval
→ candidate evidence and quality checks
→ accept, review, reject or scoped enrollment
The embedding and vector index produce an ordered list. They do not know whether the correct response is absent. Your decision layer applies a policy validated for the particular model revision, crop pipeline, gallery and cost of error.
A minimal decision record might look like this; the example score is illustrative, not a recommended cutoff:
{
"query_id": "q_1842",
"model_revision": "reid-v4",
"gallery_revision": "event-2026-07-r3",
"top_candidate": "participant_071",
"top_score": 0.74,
"quality_state": "review",
"decision": "human_review"
}
Keep the ranked candidates even when the final state is reject; they help diagnose hard negatives without turning those candidates into accepted identities.
The gallery changes the meaning of every score
Gallery design changes score behavior. Decide:
- whether one identity has one prototype or several observations;
- whether embeddings are averaged, medoids are selected, or images remain separate;
- how many sessions and viewpoints are represented;
- whether clothing changes are inside the supported scope;
- how stale observations are retired;
- how model upgrades trigger gallery regeneration;
- how corrections and deletion propagate.
One perfect-looking enrollment image can be brittle under viewpoint or occlusion changes. Conversely, adding every low-quality crop can spread an identity representation toward other people. Quality-filtered, diverse observations are usually more defensible than either extreme, but the exact aggregation rule needs evaluation.
Do not mix embeddings from incompatible model or preprocessing revisions. Distance values only have meaning inside the vector space and normalization policy that produced them.
Calibrate with the unknowns you are likely to meet
A threshold selected only from genuine pairs answers the wrong question. Open-set calibration needs both:
- known probes: queries whose identity is enrolled, with separate gallery observations;
- unknown probes: queries from identities absent from the gallery.
Unknown probes should include realistic hard negatives: similar uniforms, comparable body shapes, the same venue, partial occlusion and background patterns that can tempt the model. Random people from an unrelated image collection may make rejection look easier than production.
The decision can combine several signals:
- top candidate similarity or distance;
- gap between the first and later candidates;
- agreement across multiple query crops;
- input-quality state;
- gallery support across multiple observations;
- allowed camera, session or time context;
- explicit human review for ambiguous cases.
None of these is universally reliable. A large top-1/top-2 gap can occur when every gallery candidate is wrong. A high score can reflect similar clothing. Context can be useful but must not silently override visual evidence.
Measure acceptance and identification together
An open-set test needs three groups:
- gallery identities;
- known probes belonging to those gallery identities;
- unknown probes belonging to identities excluded from the gallery.
The 2014 open-set ReID paper used Detection and Identification Rate (DIR) with False Accept Rate (FAR). In plain language:
- DIR at rank k asks how often an enrolled query is accepted and correctly retrieved within the first (k) positions;
- FAR asks how often a query from an unenrolled identity is incorrectly accepted as somebody in the gallery.
Report them together at documented operating points. Rank-1 on known probes alone says nothing about rejection. A system can rank enrolled identities well and still accept most unknowns.
Use identity-disjoint development and final test sets, and choose the operating rule on calibration data only. Test with the expected gallery size because the chance of encountering a convincing impostor generally changes as more candidates are added.
Four outcomes, not one accuracy number
| Query state | System result | Meaning |
|---|---|---|
| Enrolled | Correct candidate accepted | Supported retrieval |
| Enrolled | Rejected or wrong candidate | Missed target |
| Not enrolled | Rejected | Correct unknown handling |
| Not enrolled | Candidate accepted | False acceptance; often the highest-risk error |
A review queue creates additional outcomes, so measure reviewer workload and correction behavior. “Sent to review” should not be counted as either automatically correct or automatically wrong.
The Adversarial Open-World Person Re-Identification paper highlights the difficult case of non-target people who look similar to enrolled targets. Its surveillance watch-list framing is not a universal product template, but the technical warning applies broadly: convincing impostors, not random negatives, define the rejection challenge.
The shortcuts that turn “unknown” into a false match
Always returning top-1. Nearest means least distant among available candidates, not correct.
Using classifier confidence as open-set probability. A model can be confident on unfamiliar input. The broader open-world review Learning and the Unknown explains why uncertainty and unknownness are not interchangeable.
Calibrating on the final test. This turns reported performance into a tuned estimate.
Using only easy unknowns. Rejection must survive same-domain lookalikes.
Collapsing unknowns together. Rejected queries do not automatically share identity.
Auto-enrolling every rejection. One person can create multiple fragments, while a bad crop can create a polluted gallery entry.
Ignoring crop failures. A wrong-person or multi-person crop can produce a plausible vector and defeat even a well-calibrated policy.
What published metrics leave unanswered
The cited research defines open-set protocols and demonstrates particular methods on particular datasets. It does not supply a universal similarity threshold, a calibrated probability that a candidate is the same person, or a guarantee that rejection will transfer to a new gallery.
The internal DINO-ReID report publishes no target-absent probe protocol, DIR/FAR curve or open-set calibration tied to an immutable serving revision. Its private Rank-1 and mAP results therefore do not prove unknown-identity rejection or production API performance. Those claims require representative unknown probes, separate calibration data and an evaluation of the exact deployed pipeline.
A review-oriented product pattern
For a review-oriented photo workflow:
- Validate that the query contains one usable person.
- Retrieve a small candidate list from the project-scoped gallery.
- Apply the versioned open-set policy.
- Auto-reject clearly unsupported results.
- Send ambiguous results to a reviewer with source images visible.
- Require explicit confirmation before merging identities.
- Record corrections without automatically using them as training truth.
- Reevaluate after gallery growth, model changes or new event domains.
Open-set Person ReID is not solved by adding an unknown label to a closed-set interface. It requires absent-identity probes, a reject-capable decision layer, realistic impostors, versioned galleries and a product that treats “no supported match” as a valid result.