{"id":"f69a4082-dd3d-4d69-a117-9bb745a19a1d","arxiv_id":"2412.08231","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A camera-aware dynamic clustering framework improves unsupervised visible-infrared person re-identification, achieving strong results on SYSU-MM01 and RegDB and narrowing the gap with supervised methods.","lead":"This paper proposes DMIC, an unsupervised method for matching people across visible and infrared cameras by making clustering aware of both modality and camera differences. Generalists might read it to see whether hand-crafted clustering tweaks still meaningfully improve person re-identification, an area where unsupervised methods are approaching supervised accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DNC's exponential schedulers (Eqs. 8-10) are ambiguous and self-referential as written; without code or an explicit per-epoch update rule, the core mechanism behind the claimed SOTA gains is not reproducible.","rationale":"DNC's dynamic schedulers are the load-bearing novelty of the paper. The reader's weakest_assumption correctly identifies that the nine scheduler hyperparameters are set per dataset without a validation protocol. My check finds a prior problem: even with the hyperparameters fixed, the equations do not specify a unique schedule. I therefore partially agree. I do not reject the paper: the central idea (camera/modality-aware distance expansion plus dynamic clustering) is plausible, and the RegDB gains are large enough that they could be real. But the missing code and ambiguous equations make the reported numbers currently unverifiable. The appropriate disposition remains conditional acceptance pending code release and an explicit schedule definition.","tokens_in":20857,"tokens_out":6548,"duration_ms":67653,"concrete_test":"Request the official code or an explicit table of eps(t), k2(t) for every epoch on both datasets; then rerun the Table IV ablation (VC+HMCL vs VC+MIE+DNC+HMCL) on RegDB using that schedule. In parallel, implement the two literal readings of Eqs. 8-10: (a) eps(e)=π2·σ_n^e in phase 1 and eps(e)=π1·σ_b^e in phase 2, and (b) interpolation between π1 and π2 over the scheduled epochs. If neither reading reproduces the reported 86.31 Rank-1 / 81.36 mAP within a few points, the paper's attribution of the gain to the described DNC mechanism is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eqs. 8 and 9 of Sec. III-C define π1 = π2·σ_n^epochs and π2 = π1·σ_b^epochs. With the reported settings (σ_n=0.5, σ_b=2 on SYSU-MM01; σ_n=2/3, σ_b=1.5 on RegDB), σ_b = 1/σ_n, so the two equations are just rearrangements of one another; neither determines a schedule for both bounds. A literal geometric-decay reading drives eps to 0.6·0.5^50 ≈ 5·10^-16 by the end of the 50-epoch intra-modality phase, far below the stated lower bound π1=0.3. A two-phase reading (shrink then expand) is described in words and Fig. 4, but no update equation or per-phase epoch indexing is given. Since DNC is the component credited for refining pseudo labels, and its ablation gain (Table IV rows 2 vs 4) is central to the paper's claim, the reported RegDB Rank-1 86.31 and SYSU-MM01 Rank-1 65.24 cannot be reproduced from the manuscript alone. This is compounded by the absence of code, error bars, or a validation protocol for the nine per-dataset DNC hyperparameters; the possibility that the gain comes from an implicit, unstated schedule or from test-set selection is therefore not excluded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DMIC, an unsupervised visible-infrared person re-identification method built from three components: Modality-Camera Invariant Expansion (MIE), which modifies k-reciprocal distance encoding by averaging over cameras; Dynamic Neighborhood Clustering (DNC), which schedules the DBSCAN radius eps and the expansion neighbor count k2 over two training phases; and Hybrid Modality Contrastive Learning (HMCL), combining intra- and inter-modality cluster/instance-level losses. The method is evaluated on SYSU-MM01 and RegDB, reporting Rank-1 of 86.31% on RegDB (vs 78.28% for DCCL) and 65.24% on SYSU-MM01 under batch size 128, and the authors claim this significantly reduces the gap to supervised methods.","tokens_in":21220,"tokens_out":6207,"duration_ms":63608,"significance":"If the reported results are reproducible, the contribution is practically relevant: the method addresses a known failure mode of clustering-based USL-VI-ReID, namely identity splitting caused by cross-camera and cross-modality variation, and the ablation tables consistently show each component adding measurable improvement. The paper also provides useful diagnostics such as cluster-count evolution and t-SNE visualizations. However, the central quantitative claims currently rest on an under-specified scheduler, per-dataset hyperparameter choices, single-run results without error bars, and an inconsistency between the text and Table I regarding which batch size supports the SOTA claim. The idea is credible and the components are clearly motivated, but the evidence as presented is not yet at the level required to certify the headline numbers.","major_comments":[{"comment":"The DNC scheduler equations are self-referential as written. Eq. (8) sets π1 = π2·σ_n^epochs and Eq. (9) sets π2 = π1·σ_b^epochs; for the reported settings (SYSU-MM01: σ_n=0.5, σ_b=2; RegDB: σ_n=2/3, σ_b=1.5), σ_b = 1/σ_n, so the two equations are algebraically identical and determine neither bound independently. A literal geometric-decay reading gives eps = 0.6·0.5^50 ≈ 5×10^-16 after 50 epochs, far below the stated lower bound π1=0.3. The two-phase interpretation (shrink, then expand) appears in the text and Fig. 4, but no piecewise formula or per-phase epoch indexing is provided. Since DNC is the component credited with refining pseudo labels and its ablation gain is central to the paper, the reported Rank-1 86.31% and 65.24% cannot be reproduced from the manuscript alone. Please provide an explicit per-epoch update rule for eps and k2 for each of the two training phases.","section":"III-C, Eqs. (8)-(10) and IV-B"},{"comment":"The text claims, \"we achieve SOTA performance ... we outperform GUR by 2.06%/1.3% in Rank-1 and 0.38%/0.95% in mAP under both testing modes.\" In Table I, DMIC(ours) with batch size 64 has all-search Rank-1 60.26% and mAP 57.82%, which is below GUR (63.51%/61.63%) and DCCL (63.18%/58.62%); the margin over GUR exists only for DMIC† with batch size 128 (65.24%/62.01%). On RegDB, however, batch size 128 degrades performance relative to batch size 64 (Rank-1 83.97% vs 86.31%). The paper therefore does not identify a single primary configuration, and the headline SOTA claim depends on a non-default setting whose effect is opposite on the two datasets. Please clarify which configuration is the primary result and discuss the batch-size sensitivity.","section":"IV-C, Table I"},{"comment":"The manuscript reports no error bars, no multiple seeds, and no validation/sensitivity analysis for the hyperparameters that define DNC. Section IV-B lists nine dataset-specific values (π2, π1, ε3, ε2, ε1, σ_n, σ_b, σ_k, k1), and the ablation in Table IV attributes only about 2-5% mAP to DNC. Given that single runs of ReID training can easily vary by more than one Rank-1 point, the claimed margin over GUR on SYSU-MM01 (1.73 points at batch 128) is not established as real. In addition, since the scheduler bounds and rates are hand-tuned per dataset, a sensitivity study (e.g., varying σ_n, π1, π2 around the chosen values) is necessary to show the method is robust rather than tuned to the two test sets. At minimum, report mean and standard deviation over several seeds.","section":"IV-B, IV-D"},{"comment":"The MIE per-camera expansion is undefined when a camera has no representative among the k2 reciprocal neighbors: Eq. (7) divides by n_c^i, which can be zero, and the text does not specify how missing cameras are handled in the average over nc cameras in Eq. (6). Since k2 can be as small as ε1=6, zero-count cameras will occur in practice, especially for the infrared cameras or minority cameras on SYSU-MM01. Please state the exact fallback used (e.g., skip absent cameras and renormalize, or compute over cameras present in the reciprocal neighbor set).","section":"III-B, Eqs. (6)-(7)"}],"minor_comments":[{"comment":"There are several typographical errors: \"stragegy\" should be \"strategy,\" \"schedular\" should be \"scheduler,\" \"unlabled\" should be \"unlabeled,\" and \"oberve\" should be \"observe.\"","section":"III-C"},{"comment":"The InfoNCE denominator in Eq. (13) sums over k=0 to I, but cluster indices appear to start at 1; the positive cluster is also included in the denominator, which is non-standard. Please clarify the indexing and whether the positive term is excluded.","section":"III-D, Eq. (13)"},{"comment":"The checkmark format in the ablation tables is ambiguous because the column headers do not align with the row entries after typesetting. Please add an explicit legend or a textual description of which components are present in each row.","section":"IV-D, Tables III and IV"},{"comment":"In the supervised section of Table I, the row labeled \"MPANet [24]\" cites reference [24], but reference [24] in the bibliography is the Channel Augmented Joint Learning paper; the MPANet attribution appears to be a citation error.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eqs. (8)-(10) is valid and should be the primary request for revision: the two scheduler equations are mutually redundant with the reported settings, so the method's key mechanism is not reproducible. I would also ask for code or a precise algorithmic description of the scheduler, and for multiple-seed results, before accepting the SOTA claims. The inconsistency between the batch-64 and batch-128 results in Table I and the text should be resolved editorially as well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is MIE: expanding the k-reciprocal distance encoding by camera-normalized averaging is a small, sensible change that addresses real cross-camera splitting in VI-ReID clustering, and the ablations give it credible credit. The RegDB result (86.31 Rank-1, up from 78.28 for DCCL) is the kind of jump the field would care about. The two-phase training idea—discriminability first, then cross-modal generalization—is reasonable.\n\nBut the DNC scheduler, the component credited for much of the gain, is not reproducible as written. Eqs. (8)-(10) are self-referential: with the reported settings, σb = 1/σn, so (8) and (9) are just rearrangements of each other and define no schedule for both bounds. A literal geometric-decay reading drives eps to ~5e-16 by epoch 50, far below the stated lower bound. The shrink-then-expand behavior is described in words and Fig. 4 but never written as an actual update rule. This is the load-bearing mechanism of the paper, so the numbers cannot be re-derived from the manuscript alone. No code, no error bars, and nine per-dataset DNC hyperparameters with no validation protocol further weaken the exact SOTA claims.\n\nThere are also smaller inconsistencies: the text says DMIC improves over GUR by 2.06% Rank-1, but the table shows that gap is actually against DCCL at batch 128; the batch-64 DMIC is below both. The claim that DNC adds \"no additional parameters\" is misleading—it adds a schedule with several hyperparameters.\n\nNone of this kills the idea. MIE is a genuine, if modest, contribution, and the ablation structure is honest. What is missing is a precise description of the schedule and a way to reproduce or at least validate the tuning. If the authors can write out the per-epoch update rule and release code, the paper deserves a serious look.\n\nI would send it to review, but I'd make code and a clear schedule a condition. For my own work, I wouldn't cite the results until the schedule is pinned down.","headline":"MIE is a sensible small idea and the ablations are credible, but the DNC scheduler is under-specified to the point of non-reproducibility, so the headline numbers need code and a clear update rule before they can be trusted.","tokens_in":21725,"tokens_out":3044,"would_cite":false,"duration_ms":30865,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that camera-aware distance expansion and a dynamic clustering schedule push unsupervised visible-infrared person re-identification to 86.31% Rank-1 on RegDB and 65.24% on SYSU-MM01, narrowing the gap to supervised methods.","keywords":["unsupervised visible-infrared person re-identification","cross-modality clustering","camera discrepancy","pseudo-label generation","dynamic neighborhood clustering","contrastive learning","DBSCAN","person re-identification"],"falsifier":"Run the full DMIC training on SYSU-MM01 using the scheduler constants reported for RegDB, with no other changes; if Rank-1 falls back to the level of the vanilla clustering baseline (around 42%), the dynamic schedule is not transferring and the central claim is not general. Alternatively, tune the scheduler constants on a held-out subset of identities and report the held-out Rank-1; if the gain over DCCL mostly vanishes, the reported test-set numbers are inflated by selection.","tokens_in":20661,"feed_emoji":"🎥","tokens_out":5069,"duration_ms":48987,"temperature":0.7,"pith_summary":"This paper tries to establish that the main remaining obstacle in unsupervised visible-infrared person re-identification is not cross-modality association but identity splitting caused by camera differences, and that this splitting can be removed by changing how distances are computed before clustering. It proposes DMIC, a training framework whose clustering step uses camera-aware distance expansion and a dynamically adjusted search radius and neighborhood size. On RegDB the method raises unsupervised Rank-1 from 78.28% to 86.31%, and on SYSU-MM01 with batch size 128 it reaches 65.24% Rank-1, both above prior unsupervised state of the art. If the result holds, clustering-based pseudo-label pipelines for cross-modal Re-ID can be improved without extra parameters or larger batches, making 24-hour person search more practical without manual identity labels.","feed_headline":"Unsupervised VI-ReID hits 86.31% Rank-1 on RegDB","feed_subtitle":"A camera-aware, dynamically scheduled clustering step cuts identity splitting and nears supervised accuracy without labels.","key_machinery":"The load-bearing object is the expanded distance encoding used as the input to DBSCAN. MIE replaces the standard k-reciprocal encoding expansion of Eq. (3) with a per-camera averaged version (Eqs. (6)-(7)), so that the embedding for each instance receives equal contribution from every camera before Jaccard distances are computed. DNC then controls two clustering hyperparameters, the DBSCAN radius eps and the expansion neighborhood size k2, through exponential schedulers (Eqs. (8)-(10)): eps shrinks to purify clusters early and grows later to merge cross-camera and cross-modality samples, while k2 follows the opposite path to gradually widen affinities. These two mechanisms change the pseudo-label generation itself rather than adding a loss term, and HMCL turns the resulting labels into training signal through cluster-level and instance-level contrastive memories.","core_discovery":"The central claim is that cross-camera variation, not just cross-modality variation, is what breaks clustering in USL-VI-ReID, and that a clustering procedure which explicitly balances distance evidence across cameras and modalities produces pseudo-labels clean enough for contrastive learning to nearly close the gap to supervised training. The paper argues that this is achieved by Modality-Camera Invariant Expansion (MIE), which averages k-reciprocal distance encodings separately per camera before fusing them, by Dynamic Neighborhood Clustering (DNC), which shrinks then expands the DBSCAN radius and the neighborhood size on an exponential schedule, and by Hybrid Modality Contrastive Learning (HMCL), which trains with intra-modality and inter-modality cluster- and instance-level losses. The reported evidence is a large improvement in Rank-1 and mAP over prior unsupervised methods on RegDB and SYSU-MM01, together with ablations showing MIE contributes roughly 10-15% mAP and DNC a further 2-5%.","pith_inferences":["If the mechanism is as general as the paper suggests, the camera-balanced distance expansion should also help single-modality unsupervised Re-ID, where camera splitting is a known failure mode but is currently handled by proxy losses rather than by changing the clustering input.","A direct transfer test would be to fix the scheduler constants (pi1, pi2, epsilons, sigmas) on RegDB and run SYSU-MM01 without re-tuning; the paper reports per-dataset values, so its claim that the schedule is parameter-free requires this to still work.","The exponential schedule has only three rates and three bounds, so a small grid search on a held-out validation split could determine how much of the gain is due to the dynamic mechanism itself versus to choosing favorable endpoints.","Future work could replace the hand-set schedule with an adaptive criterion based on cluster purity or inter-cluster stability, which would test whether the specific exponential form is essential."],"forward_implications":["Clustering-based USL-VI-ReID can treat camera identity as first-class information at the distance level instead of adding camera proxy losses.","The dynamic schedule lets a single global DBSCAN pass replace per-camera clustering while still controlling identity splitting.","The reported RegDB results imply unsupervised cross-modal Re-ID can reach performance comparable to several supervised methods without any identity annotations.","Because MIE and DNC change only the clustering input and hyperparameters, they can be dropped into other unsupervised Re-ID frameworks that rely on DBSCAN pseudo-labels."],"supporting_citations":[{"why":"Supplies the DBSCAN density-based clustering algorithm that assigns pseudo-labels from the distance embeddings.","marker":"[12]"},{"why":"Supplies the k-reciprocal distance encoding that MIE modifies with per-camera averaging.","marker":"[39]"},{"why":"Supplies the AGW backbone and the evaluation metrics used for all experiments.","marker":"[3]"},{"why":"Supplies the Channel exchangeable Augmentation that generates the extra visible-stream features used in training and memory updates.","marker":"[24]"},{"why":"GUR is the prior unsupervised method that clusters within cameras and uses cross-camera proxies, the main comparison for the global clustering approach.","marker":"[37]"},{"why":"DCCL is the strongest unsupervised baseline on RegDB whose Rank-1 of 78.28% is raised to 86.31%.","marker":"[36]"},{"why":"Provides the real-time memory updating idea that HMCL adapts with randomly selected samples.","marker":"[21]"},{"why":"Defines the SYSU-MM01 benchmark and its all-search and indoor-search evaluation settings.","marker":"[41]"},{"why":"Defines the RegDB benchmark and its visible-to-infrared and infrared-to-visible evaluation settings.","marker":"[42]"}],"fun_headline_variants":["Cross-camera variance splits identities in unsupervised VI-ReID","Dynamic camera-aware clustering nears supervised VI-ReID accuracy","DMIC: modality-camera invariant clustering reduces identity splits","Unsupervised VI-ReID: camera-aware expansion improves pseudo-labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-set schedules for the clustering radius and neighborhood size, with different starting values chosen for each dataset, are what produce the reported gains; the paper does not describe how these values were selected, so if they were tuned on the test sets the improvement would partly reflect parameter selection.","fun_headline_variants_meta":{"raw":{"variants":["Cross-camera variance splits identities in unsupervised VI-ReID","Dynamic camera-aware clustering nears supervised VI-ReID accuracy","DMIC: modality-camera invariant clustering reduces identity splits","Unsupervised VI-ReID: camera-aware expansion improves pseudo-labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1786,"prompt_tokens":1052,"completion_tokens":734,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":664}},"tokens_in":668,"tokens_out":734,"duration_ms":7918,"temperature":1.0,"reasoning_tokens":664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:02:56.643355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full DMIC training on SYSU-MM01 using the scheduler constants reported for RegDB, with no other changes; if Rank-1 falls back to the level of the vanilla clustering baseline (around 42%), the dynamic schedule is not transferring and the central claim is not general. Alternatively, tune the scheduler constants on a held-out subset of identities and report the held-out Rank-1; if the gain over DCCL mostly vanishes, the reported test-set numbers are inflated by selection.","supporting_citations":[{"cited_title":"Re-ranking person re- identification with k-reciprocal encoding,","cited_arxiv_id":null,"evidence_quote":"Supplies the k-reciprocal distance encoding that MIE modifies with per-camera averaging."},{"cited_title":"Deep learning for person re-identification: A survey and outlook,","cited_arxiv_id":null,"evidence_quote":"Supplies the AGW backbone and the evaluation metrics used for all experiments."},{"cited_title":"Ester, H","cited_arxiv_id":null,"evidence_quote":"Supplies the DBSCAN density-based clustering algorithm that assigns pseudo-labels from the distance embeddings."},{"cited_title":"Channel augmented joint learning for visible-infrared recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the Channel exchangeable Augmentation that generates the extra visible-stream features used in training and memory updates."},{"cited_title":"Towards grand unified representation learning for unsupervised visible-infrared person re- identification,","cited_arxiv_id":null,"evidence_quote":"GUR is the prior unsupervised method that clusters within cameras and uses cross-camera proxies, the main comparison for the global clustering approach."},{"cited_title":"Dual Consistency-Constrained Learning for Unsupervised Visible-Infrared Person Re-Identification,","cited_arxiv_id":null,"evidence_quote":"DCCL is the strongest unsupervised baseline on RegDB whose Rank-1 of 78.28% is raised to 86.31%."},{"cited_title":"A real-time memory updating strategy for unsupervised person re-identification,","cited_arxiv_id":null,"evidence_quote":"Provides the real-time memory updating idea that HMCL adapts with randomly selected samples."},{"cited_title":"Rgb-infrared cross-modality person re-identification,","cited_arxiv_id":null,"evidence_quote":"Defines the SYSU-MM01 benchmark and its all-search and indoor-search evaluation settings."},{"cited_title":"Person recognition system based on a combination of body images from visible light and thermal cameras,","cited_arxiv_id":null,"evidence_quote":"Defines the RegDB benchmark and its visible-to-infrared and infrared-to-visible evaluation settings."}],"review_version":1}