Anytime Person ReID: Multi-Scenario Retrieval

• AT-ReID is a person re-identification framework that retrieves targets across diverse scenarios (day/night, short- and long-term) by jointly addressing modality, temporal, and clothing challenges.
• Uni-AT and STFER exemplify scenario-aware modeling, where dedicated modules like CLS tokens, Mixture-of-Attribute-Experts, and semantic-driven token filtering enhance identity representation.
• The AT-USTC benchmark, featuring over 400K images across six distinct conditions, underpins robust training and cross-dataset generalization for AT-ReID systems.

Searching arXiv for the cited AT-ReID and related papers to ground the article in the latest available metadata. Anytime Person Re-identification (AT-ReID) is a person re-identification setting in which a system must retrieve a target person at any time, including daytime and nighttime, and across short-term and long-term intervals. In the formulation introduced with AT-USTC and Uni-AT, the task is decomposed into six non-overlapping scenarios—DT-ST, DT-LT, NT-ST, NT-LT, AD-ST, and AD-LT—covering RGB, infrared, cross-modality, and clothing-change conditions within a single framework (Li et al., 20 Sep 2025). Subsequent work argues that AT-ReID is fundamentally harder than traditional ReID, cross-modality ReID, or cloth-changing ReID in isolation because it must jointly address modality shifts, extensive clothing changes, and temporal variation, while retaining discriminative identity features under scenario-specific biases (Li et al., 16 Apr 2026).

1. Definition and task scope

AT-ReID extends standard person ReID from a single-scenario retrieval problem to a multi-scenario setting. In general ReID, given a query image $x_q$ and gallery set $\mathcal{G}=\{x_g\}$, the aim is to learn an embedding function $f(\cdot)$ such that same-identity samples are close in feature space and different identities are far apart: $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$ AT-ReID retains this retrieval objective but requires the learned representation to remain valid under six scenarios: Daytime Short-Term (DT-ST), Daytime Long-Term (DT-LT), Nighttime Short-Term (NT-ST), Nighttime Long-Term (NT-LT), All-day Short-Term (AD-ST), and All-day Long-Term (AD-LT) (Li et al., 16 Apr 2026).

In this setting, “anytime” denotes arbitrary temporal intervals, from minutes or hours to days or weeks, together with day-night shifts, RGB–IR cross-modality, and clothing change. The six-scenario formulation organizes these factors through time moment attributes—DT, NT, AD—and time interval attributes—ST, LT (Li et al., 20 Sep 2025). The resulting problem is not merely long-term ReID: it is a joint long-term, cross-modality, and cloth-changing retrieval task.

The distinction from earlier ReID subfields is structural. Traditional ReID is typically RGB-only and short-term; cross-modality ReID addresses RGB–IR matching under clothing-consistent conditions; cloth-changing ReID focuses on clothing variation, usually within a single modality; and illumination-adaptive ReID isolates illumination as the primary varying factor across time (Li et al., 20 Sep 2025). AT-ReID generalizes these settings by requiring one model to operate across all such conditions simultaneously. This suggests that feature salience is scenario-dependent: clothing is strongly discriminative in short-term RGB retrieval, but harmful in long-term or cross-modality retrieval, while RGB-specific color cues are useful in DT-ST yet detrimental in AD-ST or AD-LT.

2. Benchmark construction and AT-USTC

AT-USTC is the first dedicated benchmark for AT-ReID. It contains 403,599 images of 270 identities, including 199,803 RGB images and 203,796 IR images, acquired by 16 non-overlapping cameras—8 RGB and 8 IR—across 5 indoor and 11 outdoor scenes (Li et al., 20 Sep 2025). The dataset spans 21 months and includes 710 different clothing sets, with 2–14 outfits per training identity and an average of 3.6 outfits. Each volunteer was photographed 29.1 times on average across different dates or scenes, and identities appear in 5.7 RGB and 5.3 IR cameras on average.

These collection choices are central to the benchmark’s role. Existing datasets generally omit at least one of long-term clothing change, infrared modality, or repeated capture over long temporal horizons. AT-USTC was designed to provide conditions for all six scenarios, including NT-LT and AD-LT, which were largely unsupported in prior benchmarks (Li et al., 20 Sep 2025). The dataset therefore operationalizes AT-ReID as a unified evaluation problem rather than a loose aggregation of existing tasks.

The standard split uses 135 identities for training and 135 for testing. The training partition contains 286,087 images, with 20% of them—55,060 images—used for validation, while the testing partition contains 117,512 images (Li et al., 16 Apr 2026). Evaluation follows scenario-specific query/gallery splits and reports Rank-$k$ accuracy and mean Average Precision (mAP), with “Any-Time” defined as the average performance over all six scenarios. This average is the principal summary metric because it measures the ability to retrieve under any scenario rather than optimize for one.

A further implication of AT-USTC is its value as a source dataset for generalization. With a standard ResNet50, training on AT-USTC yields higher average Rank-1 and mAP across Market1501, CUHK03, SYSU-MM01, PRCC, and LTCC than training on MSMT17, LLCM, or DeepChange (Li et al., 20 Sep 2025). This suggests that high intra-identity diversity and joint scenario coverage produce representations that transfer more effectively beyond the native AT-ReID benchmark.

3. Scenario-aware modeling before semantic guidance

The first dedicated model for the benchmark is Uni-AT, a unified architecture built around a ViT-Base backbone, six scenario CLS tokens, a Mixture-of-Attribute-Experts (MoAE) module, and a Hierarchical Dynamic Weighting (HDW) strategy (Li et al., 20 Sep 2025). Rather than using one global CLS token, Uni-AT assigns one CLS token $t_i^s$ to each scenario $s$, allowing the shared patch-token representation to be queried in six different ways.

Its training objective is a scenario-aware identity loss: $$\mathcal{L}^s_{id}(t_i^s) = -\log(p^s_{i,gt}), \qquad p^s_{i,gt} = \frac{\exp(o^s_{i,gt})}{\exp(o^s_{i,gt}) + \sum_{j \in N_s} \exp(o^s_{i,j})}.$$ The definition of the negative category set $N_s$ depends on the scenario. In short-term scenarios, categories are defined by clothes IDs, but only for clothing sets whose owners differ from the ground-truth clothes’ owner. In long-term scenarios, categories are person IDs (Li et al., 20 Sep 2025). This design formalizes a central AT-ReID asymmetry: clothing is useful in ST but must be discounted in LT.

Uni-AT also filters supervision by modality. RGB images supervise DT-ST, DT-LT, AD-ST, and AD-LT tokens, while IR images supervise NT-ST, NT-LT, AD-ST, and AD-LT tokens (Li et al., 20 Sep 2025). This avoids learning DT-specific classifiers from nighttime infrared data or NT-specific classifiers from daytime RGB imagery.

MoAE is intended to reduce inter-scenario interference in the shared ViT. For scenario $s$, experts are selected by sparse routing,

$\mathcal{G}=\{x_g\}$0

with $\mathcal{G}=\{x_g\}$1. Experts are composed from five attribute layers: DT, AD, NT for time moment, and ST, LT for time interval. An expert for a scenario is built by combining one time-moment layer and one time-interval layer: $\mathcal{G}=\{x_g\}$2 HDW then weights scenario losses hierarchically,

$\mathcal{G}=\{x_g\}$3

with

$\mathcal{G}=\{x_g\}$4

The attribute confidences are averaged over all scenarios sharing the same time moment or time interval (Li et al., 20 Sep 2025). This converts AT-ReID into a structured multi-task problem in which related scenarios influence one another through explicit attribute groupings.

Empirically, Uni-AT achieves 55.80 Rank-1 and 41.38 mAP on the Any-Time metric, improving over baseline MS-ReID at 50.90 and 34.75, and over task-specific methods such as CLIP-ReID (ViT-B), CCIL, and DEEN on the benchmark average (Li et al., 20 Sep 2025). These results establish the scenario-aware formulation as the first effective baseline for AT-ReID, while also exposing its remaining limitations under AD-LT and NT-LT.

4. Semantic-driven Token Filtering and Expert Routing

Semantic-driven Token Filtering and Expert Routing (STFER) introduces a semantically guided cross-modal framework specifically for AT-ReID and frames its contribution as moving beyond pure visual features (Li et al., 16 Apr 2026). The method combines a ViT backbone, an offline Large Vision-LLM (LVLM), Semantic-driven Visual Token Filtering (SVTF), and Semantic-driven Expert Routing (SER).

The offline LVLM is Qwen3-VL-4B. For each identity $\mathcal{G}=\{x_g\}$5, the method randomly samples $\mathcal{G}=\{x_g\}$6 images across clothing and modality conditions and uses instructions to generate identity-consistent textual descriptions: $\mathcal{G}=\{x_g\}$7

$\mathcal{G}=\{x_g\}$8

The descriptions are intended to encode intrinsic attribute descriptions such as body type, gender presentation, age range, hair style or color, and facial attributes, while avoiding transient clothes (Li et al., 16 Apr 2026). This suggests a semantic abstraction layer that is less sensitive to modality-specific appearance and outfit variation than raw pixels.

Text and image features are then embedded as

$\mathcal{G}=\{x_g\}$9

and concatenated with six scenario CLS tokens: $f(\cdot)$0 This gives text tokens a direct role in ViT processing rather than treating them as external supervision (Li et al., 16 Apr 2026).

SVTF uses text tokens as queries and visual tokens as keys: $f(\cdot)$1

$f(\cdot)$2

The refined visual tokens are written as

$f(\cdot)$3

Conceptually, the module enhances patches aligned with identity-intrinsic semantics and suppresses background, clothing-dependent patterns, and other redundancies. Visualization reported for STFER shows attention concentrating on body parts rather than background or clothing (Li et al., 16 Apr 2026).

SER extends earlier mixture-of-experts ideas by conditioning routing on both scenario-specific CLS features and global text embeddings. With

$f(\cdot)$4

stochastic text masking is applied: $f(\cdot)$5 with $f(\cdot)$6. The routing weights are then

$f(\cdot)$7

and the top-$f(\cdot)$8 experts are aggregated as

$f(\cdot)$9

This yields a scenario-specific identity representation informed by semantic identity text as well as visual context (Li et al., 16 Apr 2026).

The training objective remains scenario-aware identity classification: $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$0 No explicit triplet, contrastive, or alignment losses are described; semantic alignment is instead embedded in SVTF and SER through token interaction and routing (Li et al., 16 Apr 2026).

5. Relation to earlier robustness paradigms

AT-ReID emerged from limitations already identified in earlier subproblems. Apparel-changed ReID argued that most public datasets were collected in short time windows and that deep ReID models are heavily biased to apparel and color features, which becomes problematic when the same person changes clothes or different people wear similar clothes (Yu et al., 2020). Apparel-invariant Feature Learning (AIFL) addressed this by minimizing the distance between samples of the same person under different apparel: $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$1 using reconstruction and adversarial objectives on synthetic cloth-changed pairs. Its joint objective is

$$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$2

with $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$3 (Yu et al., 2020). This line of work directly anticipates the long-term component of AT-ReID by treating clothing as a nuisance factor for identity retrieval.

Illumination-Adaptive ReID treated long-term retrieval as a problem of matching across different illuminations at different times across a day and introduced Illumination-Identity Disentanglement (IID) to separate identity-relevant and identity-irrelevant illumination factors (Zeng et al., 2019). With

$$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$4

and reconstructed image

$$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$5

the training objective becomes

$$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$6

with typical settings $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$7 and $$z_q = f(x_q),\quad z_g = f(x_g), \quad \text{and} \quad \text{sim}(z_q, z_g^{\text{same ID}}) > \text{sim}(z_q, z_g^{\text{different ID}}).$$8 (Zeng et al., 2019). The paper characterizes illumination-adaptive ReID as a practical long-term retrieval problem and shows that disentangled identity features improve robustness to multi-illumination conditions.

Within the AT-ReID literature, these earlier approaches are best understood as partial solutions. AIFL targets apparel invariance but not RGB–IR modality shifts. IID targets illumination changes but not cloth-changing or six-scenario retrieval. AT-ReID integrates these previously separate axes into a single benchmark and model class (Li et al., 20 Sep 2025). STFER makes this integration explicit by using identity-consistent semantic text as a common anchor across clothing, time, and modality (Li et al., 16 Apr 2026). A plausible implication is that AT-ReID consolidates several robustness traditions—apparel invariance, illumination robustness, and cross-modality matching—under a scenario-aware representation learning regime.

6. Empirical performance, limitations, and outlook

On AT-USTC, STFER reports state-of-the-art performance. The previous AT-ReID result from Uni-AT is 55.80 Rank-1 and 41.38 mAP on the Any-Time metric, whereas STFER reports 94.54 Rank-1 and 93.46 mAP (Li et al., 16 Apr 2026). The paper states improvements of +69.43% in Rank-1 and +125.86% in mAP over the previous best AT-ReID method in the Any-Time scenario. Scenario-wise, STFER reports DT-ST 98.04/95.72, DT-LT 91.89/90.85, NT-ST 95.12/93.52, NT-LT 98.00/97.36, AT-ST 89.73/89.00, and AT-LT 94.46/94.29 in Rank-1/mAP. These values indicate uniformly high performance across all six scenarios, including the all-day long-term setting.

Cross-dataset generalization is also emphasized. A model trained on AT-USTC and tested zero-shot on Market1501, CUHK03, SYSU-MM01, PRCC, and LTCC obtains average performance of 74.33 Rank-1 and 75.26 mAP, compared with 42.97 Rank-1 and 31.68 mAP for the best prior method, CLIP-ReID ViT-B (Li et al., 16 Apr 2026). Per-dataset results for STFER are 93.79/90.02 on Market1501, 56.00/57.36 on CUHK03, 64.86/70.19 on SYSU-MM01, 66.95/70.25 on PRCC, and 90.05/88.50 on LTCC.

Ablation results identify the source of these gains. A visual-only baseline corresponding to a Uni-AT-style system gives 55.80 Rank-1 and 41.38 mAP on Any-Time. Adding text priors without SVTF or SER raises this to 93.45 and 92.65; adding SVTF yields 93.90 and 92.98; adding SER yields 93.79 and 92.80; and the full STFER reaches 94.54 and 93.46 (Li et al., 16 Apr 2026). The largest contribution is thus attributed to LVLM-generated semantic text, while SVTF and SER provide consistent incremental gains.

The method’s limitations are also stated. LVLM text quality weakens on low-quality images with heavy occlusion or extreme low resolution, where the model tends to produce neutral or vague descriptions. The framework assumes an identity-level text library generated from multiple images per identity, which requires offline preprocessing. There is no explicit handling of extreme occlusion, and heavy occlusion remains difficult when visual evidence for body shape or facial cues is limited (Li et al., 16 Apr 2026). In deployment terms, LVLM generation is offline, while online inference uses only ViT, SVTF, and SER.

AT-ReID research therefore occupies a transitional position in the broader ReID landscape. Uni-AT established the problem as a multi-scenario benchmark with scenario-aware learning (Li et al., 20 Sep 2025). STFER reoriented the field from pure visual cues toward semantics derived from vision-LLMs (Li et al., 16 Apr 2026). Earlier apparel-invariant and illumination-adaptive work provides the conceptual precursors for this shift (Yu et al., 2020, Zeng et al., 2019). This suggests that future AT-ReID systems may continue toward hybrid semantic–visual identity representations, richer scenario-conditioned routing, and broader multimodal formulations in which text, RGB, IR, and possibly other sensing modalities jointly support retrieval under arbitrary time and condition changes.