{"id":"62c738b2-1dec-4af8-88ea-e1c1b55c34a6","arxiv_id":"2509.16635","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AT-USTC, a 403k-image RGB/IR dataset covering six time-based ReID scenarios, and Uni-AT, a multi-scenario model, are proposed, with Uni-AT achieving 55.8% any-time Rank-1 on the new benchmark.","lead":"This paper introduces a new benchmark dataset and model for person re-identification across different times of day and time gaps, combining day and night imagery with clothing changes. It aims to push ReID systems toward realistic anytime surveillance use, but the core results currently rest on a single run with an under-specified evaluation protocol.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. 3's uniform query/gallery construction ('same camera, same clothing') contradicts the LT scenario definitions; if literal, AT-USTC's long-term scenarios do not test clothing change, undermining the benchmark's central claim.","rationale":"After reading the paper, the evaluation protocol for the six scenarios is indeed the most load-bearing unstated assumption. The paper's contributions are a new dataset and a unified model; both are only meaningful if the benchmark measures what it claims. Section 3's description of selecting images with same camera and same clothing, if applied across all scenarios, directly contradicts the definitions of DT-LT and NT-LT, which require clothing changes, and AD-ST/AD-LT, which require cross-modality. The large performance gap between DT-ST and DT-LT in the experiments (Table 3) suggests that clothing variation is present in the LT splits, so the sentence is likely a simplification, but the paper does not clarify. This ambiguity is not resolved by the experiments, the ablations, or the cross-dataset results; it is a precondition for interpreting all of them. Therefore, the reader's CONDITIONAL verdict is appropriate; no change is needed, but the authors must clarify the protocol before the benchmark is adopted. I agree with the reader's identification of this as the weakest assumption.","tokens_in":18225,"tokens_out":4558,"duration_ms":36965,"concrete_test":"Download the released AT-USTC dataset (or request access) and compute, for each identity in each of the six scenario splits, the clothing ID overlap and camera overlap between the query and gallery sets. Specifically, determine the fraction of identities where query and gallery images in DT-LT and NT-LT have identical clothing IDs. If this fraction is high (>90%), the long-term scenarios are not operationalized as clothing-change tasks, and the benchmark's coverage claim fails. If the fraction is low, the Sec. 3 sentence is misleading and must be revised; either way, the test resolves the ambiguity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that AT-USTC is the first benchmark covering all six time-based scenarios. The load-bearing premise is that the evaluation protocol correctly operationalizes those scenarios. Section 3, in describing the construction of query and gallery sets, states: 'For each identity, we selected three query images and three gallery images from video clips featuring the same identity, captured by the same camera, and with the same clothing.' Taken literally, this protocol would eliminate the very variations that define the scenarios: DT-LT and NT-LT would have identical clothing between query and gallery, so they would not be clothing-change scenarios; AD-ST and AD-LT would not require cross-modality or cross-camera matching. The experimental results undermine this literal reading—e.g., in Table 3 the baseline DT-LT Rank-1 (32.99) is far lower than DT-ST (95.02), suggesting that the LT split is genuinely harder, plausibly due to clothing variation. This internal tension leaves the actual evaluation protocol under-specified. If the sentence is a simplification (e.g., describing only the raw image selection rather than the scenario-specific task construction), the paper must state that explicitly and describe how each scenario's query/gallery sets are constructed. Without that, the benchmark's validity—and hence every reported Any-Time number—rests on an unverified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Anytime Person Re-identification (AT-ReID), a multi-scenario re-identification task spanning daytime/nighttime and short-term/long-term retrieval, and introduces AT-USTC, a 403,599-image dataset of 270 volunteers captured by 16 RGB/IR cameras over 21 months with multiple clothing per identity. The authors also propose Uni-AT, a unified model combining a Multi-Scenario ReID transformer with six scenario-specific CLS tokens, a Mixture-of-Attribute-Experts module, and Hierarchical Dynamic Weighting. Experiments on AT-USTC report Any-Time Rank-1/mAP of 55.80/41.38, outperforming task-specific baselines and existing MoE variants; cross-dataset experiments on Market1501, CUHK03, SYSU-MM01, PRCC, and LTCC are also reported.","tokens_in":18613,"tokens_out":6995,"duration_ms":63209,"significance":"If the evaluation protocol is valid, this is a substantial benchmark contribution: AT-USTC is the first large-scale dataset with RGB and IR imagery, multiple clothing changes, and long-term multi-season collection covering all six time-based scenarios. The dataset scale, privacy safeguards, and code/data release are concrete strengths, as are the cross-dataset generalization experiments and component-wise ablations. However, the central claim rests on whether the six scenarios are actually constructed as defined, and the current description is ambiguous enough to require major revision before the benchmark numbers can be interpreted.","major_comments":[{"comment":"The evaluation protocol is under-specified. The sentence \"For each identity, we selected three query images and three gallery images from video clips featuring the same identity, captured by the same camera, and with the same clothing\" would, taken literally, remove the cross-camera, cross-modality, and clothing-change variations that define DT-LT, NT-ST, NT-LT, AD-ST, and AD-LT. It also conflicts with the next sentence reporting an average gallery size of about 25 images per identity after selecting only three gallery images per identity. Table 3 suggests the LT scenarios are harder (DT-LT Rank-1 32.99 vs DT-ST 95.02), so the actual protocol likely differs from the text. Please specify, for each of the six scenarios, query and gallery modalities, camera overlap, clothing overlap, and image counts per identity; if the quoted sentence is a simplification, say so explicitly and describe th","section":"Section 3 (Data Split)"},{"comment":"The performance differences supporting the method are often small: MoAE adds 1.67 Rank-1 over the Ls_id baseline, HDW adds 1.29, and the best Uni-AT is only 2.28 Rank-1 ahead of the strongest CM-ReID baseline (DEEN) in Table 5. No error bars, multiple seeds, or statistical tests are reported. Since the central claim is that Uni-AT outperforms task-specific methods across all scenarios, single-run results are not sufficient. Please report mean and standard deviation over at least three runs, or otherwise justify that the reported gaps are larger than run-to-run variation.","section":"Section 5 (Tables 3-5)"},{"comment":"The cross-dataset evaluation is not fully defined. Existing datasets cover only subsets of the six AT-ReID scenarios (e.g., Market1501 is essentially daytime short-term, PRCC/LTCC are daytime long-term, SYSU-MM01 is all-day short-term), yet Table 6 reports a single average R1/mAP and the text refers to cross-dataset Any-Time performance. Please describe how each dataset was mapped to scenarios, which CLS token or feature was used for each test set, and what the average is over. Without this, the generalization claim cannot be interpreted.","section":"Section 5.4 / Table 6"}],"minor_comments":[{"comment":"The sentence listing cross-dataset improvements has a typo: \"10.71% / 8.64% / 11.58% and 4.72% / 6.13%\" does not match the five datasets or the values in Table 4 (the SYSU-MM01 improvement appears to be 11.48, not 11.58).","section":"Section 5.3"},{"comment":"Please reconcile \"three gallery images per identity\" with \"the gallery contains an average of approximately 25 images per identity\"; if the three images are per scenario, state that explicitly.","section":"Section 3"},{"comment":"Minor language issue: \"This puts more challenges on the ReID system\" should be \"This poses more challenges to the ReID system\".","section":"Section 1"},{"comment":"Panel (c) and (d) would benefit from axis labels and units. The text refers to \"Capture Times/Videos Per Person\" and \"Scale\", but the figures do not make clear what the plotted values are.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"I would ask the editor to request the authors to release the exact query/gallery split metadata and a precise per-scenario protocol description. The benchmark's validity is conditional on this; if the protocol turns out to be sound, the paper is likely publishable after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know is that the dataset is the real contribution, and the evaluation protocol as written has a hole that someone needs to fix before relying on any of the reported numbers.\n\nWhat's new: AT-USTC is the first large-scale benchmark that combines RGB and IR capture with long-term clothing change for the same identities—270 identities, 403k images, 21 months, 16 cameras. That covers a scenario space no existing dataset covers. The domain-generalization experiment in Table 1 is a strong piece of evidence: a plain ResNet50 trained on AT-USTC transfers better to Market1501, CUHK03, SYSU-MM01, PRCC, and LTCC than the same backbone trained on MSMT17, LLCM, or DeepChange. That point about intra-identity diversity is well made.\n\nThe method, Uni-AT, is not a conceptual breakthrough—it is a sensible combination of established MoE and multi-task weighting tricks—but the ablations are clean, each component adds a few points, and the efficiency comparison against MMoE/PLE/VLMo is fairly reported. The claim that task-specific systems fail outside their target scenario is supported by Table 5.\n\nThe soft spot is the benchmark's evaluation protocol. Section 3 says: \"For each identity, we selected three query images and three gallery images from video clips featuring the same identity, captured by the same camera, and with the same clothing.\" Taken literally, this would make the long-term scenarios contain no clothing change and the all-day scenarios no cross-modality or cross-camera matching—which guts the central claim. The numbers show the sentence can't be literal: baseline DT-LT Rank-1 is 32.99 versus 95.02 for DT-ST, so the LT split is clearly harder. But the paper never explains how the six query/gallery sets are actually built per scenario. This is load-bearing and needs a clear revision, ideally with exact query/gallery IDs released.\n\nMinor issues: no error bars or multiple seeds; the test set is only 135 identities; and the HDW weights are computed from the model's own confidences, which is acceptable as a training heuristic but not a principled reweighting.\n\nOverall, the dataset deserves serious attention and the paper belongs in the review process, but I would want the protocol clarified before treating the benchmark numbers as solid or asking anyone to build on them.","headline":"AT-USTC is a genuinely useful new ReID benchmark, but the evaluation protocol in Section 3 is written in a way that undercuts its own scenarios and needs a concrete fix before the numbers can be trusted.","tokens_in":19023,"tokens_out":3576,"would_cite":true,"duration_ms":30822,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Person re-identification can be unified into a single 'anytime' task spanning day/night and short/long time gaps, backed by a new 403k-image dataset and a multi-scenario model.","keywords":["person re-identification","anytime retrieval","cross-modality ReID","cloth-changing ReID","multi-scenario learning","mixture-of-experts","benchmark dataset","surveillance"],"falsifier":"An independent audit of AT-USTC's test splits: count, for each identity in the DT-LT and NT-LT galleries, how many distinct outfits appear across their gallery images. If every identity's long-term gallery contains only one outfit per camera, then the reported long-term numbers cannot measure cloth-changing retrieval, and the Any-Time average would be inflated by a protocol mismatch.","tokens_in":18188,"feed_emoji":"🕐","tokens_out":4405,"duration_ms":36095,"temperature":0.7,"pith_summary":"This paper introduces a new person re-identification task, Anytime Re-ID (AT-ReID): retrieving the same person whether the query and gallery images were captured during the day or night and whether the interval between captures is short or long enough for a clothing change. To support it, the authors built AT-USTC, a 403k-image dataset of 270 volunteers photographed repeatedly across 21 months with both RGB and infrared cameras, giving each identity many outfits, cameras, and sessions. They also propose Uni-AT, a single model that learns separate feature spaces for the six time-based scenarios and shares attribute-level experts among related ones. On AT-USTC, Uni-AT reaches 55.80 Rank-1 and 41.38 mAP across the six scenarios, and it transfers to existing ReID datasets better than task-specific methods. The point is that one model can replace separate daytime, nighttime, short-term, and long-term systems.","feed_headline":"One model retrieves people day or night, short or long term","feed_subtitle":"New AT-USTC dataset covers all six time-based ReID scenarios; Uni-AT beats task-specific methods on every one.","key_machinery":"Three components carry the method. Multi-Scenario ReID (MS-ReID) attaches six CLS tokens to a Vision Transformer, one per scenario, each supervised by a scenario-aware identity loss that treats clothing IDs as categories in short-term scenarios and person IDs in long-term scenarios. Mixture-of-Attribute-Experts (MoAE) routes each scenario token through a sparse expert formed by combining a time-moment attribute layer (DT, AD, or NT) with a time-interval attribute layer (ST or LT), letting related scenarios share parameters. Hierarchical Dynamic Weighting (HDW) scales each scenario loss by the average confidence of scenarios sharing the same time-moment or time-interval attribute, balancing t","core_discovery":"The paper's central claim is that the six time-based retrieval scenarios—daytime short-term, daytime long-term, all-day short-term, all-day long-term, nighttime short-term, nighttime long-term—can be unified into a single task, and that a single model can outperform methods specialized for one scenario on all six. The evidence is AT-USTC, the first dataset that contains both RGB and infrared imagery and clothing changes per identity, and Uni-AT, which uses six scenario-specific classification tokens, a Mixture-of-Attribute-Experts module that shares time-moment and time-interval layers across related scenarios, and a hierarchical dynamic weighting loss. The authors report that Uni-AT achieve","pith_inferences":["If the protocol is audited and confirmed to include clothing changes within long-term galleries, AT-USTC could become the standard benchmark for evaluating 'anytime' capability; if not, the long-term numbers would need reinterpretation.","The attribute-layer sharing idea could transfer to other multi-scenario recognition problems, such as recognizing people across camera viewpoints or domains, where scenarios share partial attributes.","Because the dataset spans 21 months, it may also enable studies of how appearance changes over time, e.g., predicting a person's current look from past captures, which the paper does not explore.","One testable extension: train Uni-AT on AT-USTC and evaluate on a held-out subset of another cloth-changing dataset to see whether the any-time gains persist when the IR modality is absent."],"forward_implications":["AT-USTC can serve as a common testbed for any model that claims to work across day/night and short/long-term retrieval, since it is the first dataset covering all six scenarios.","A single Uni-AT-style model can replace separate traditional, cross-modality, and cloth-changing ReID systems, because it outperforms task-specific baselines on its own scenarios as well as the others.","The dataset's high intra-identity diversity (29.1 captures per person) improves cross-dataset generalization, suggesting that ReID datasets should be judged not only by number of identities but by per-identity variation.","The scenario-aware identity loss—clothing categories for short-term, person IDs for long-term—lets the model exploit clothing when it is reliable and ignore it when it is not."],"fun_headline_variants":["AT-USTC: first dataset for anytime person re-identification","Uni-AT unifies day, night, short, and long-term person search","Person re-ID anytime: one model beats scenario-specific methods","Benchmark for anytime ReID covers all six time scenarios","Anytime ReID: 403k images, six scenarios, one model"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The benchmark's validity rests on the unstated assumption that the six-scenario test protocol actually contains the intended variations—especially that long-term query-gallery pairs involve clothing changes and that this holds across modalities; the paper's one-line description of query/gallery selection says images come from the same camera and same clothing, which, taken literally, would remove the very variation the long-term scenarios are meant to test.","fun_headline_variants_meta":{"raw":{"variants":["AT-USTC: first dataset for anytime person re-identification","Uni-AT unifies day, night, short, and long-term person search","Person re-ID anytime: one model beats scenario-specific methods","Benchmark for anytime ReID covers all six time scenarios","Anytime ReID: 403k images, six scenarios, one model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001044,"raw_usage":{"total_tokens":4255,"prompt_tokens":800,"completion_tokens":3455,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":3371}},"tokens_in":544,"tokens_out":3455,"duration_ms":21986,"temperature":1.0,"reasoning_tokens":3371,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:04:02.008284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent audit of AT-USTC's test splits: count, for each identity in the DT-LT and NT-LT galleries, how many distinct outfits appear across their gallery images. If every identity's long-term gallery contains only one outfit per camera, then the reported long-term numbers cannot measure cloth-changing retrieval, and the Any-Time average would be inflated by a protocol mismatch.","supporting_citations":[],"review_version":1}