{"id":"10919061-f159-4f5a-b703-ff000cb25a51","arxiv_id":"2505.02549","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RoDE, a dual-model training framework with adaptive sample weighting and cluster consistency matching, improves unsupervised visible-infrared person re-identification accuracy on SYSU-MM01, RegDB, and LLCM.","lead":"This paper presents a new unsupervised method that matches people across visible and infrared camera images without needing labeled data. The method uses two models that train each other, plus a cluster-alignment step, and reports improved accuracy on three public re-identification benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RAL loss Eq. (7) is mis-stated: -p^gamma does not tend to CE as gamma->0, and the adaptive weight gamma_i is defined self-referentially through Eq. (18), so the core robust-learning objective is not well-defined as written.","rationale":"The reader's verdict is conditional, and I agree that conditional acceptance pending mathematical correction and artifact release is the right disposition. However, the most load-bearing weak point is not the CCM residual mismatch flagged in the reader's weakest_assumption. The 15-35% mismatching rate reported in Figure 9(e/f) is an acknowledged, partially quantified limitation; a noise-robust framework can in principle tolerate residual correspondence errors, and the ablation in Table III shows CCM is necessary but does not by itself show the alignment assumption is invalid. The more serious problem is that the RAL mechanism, which is the component specifically designed to combat noise overfitting, is not correctly specified. Eq. (7) has a property claim that is contradicted by elementary calculus, and the adaptive weights in Eq. (9) are defined in terms of a per-sample loss that already includes those weights. This makes the core optimization objective ambiguous and prevents independent verification of the central claim. The CCM concern raised by the reader is real and worth testing, but the RAL formulation issue is more fundamental because it affects the definition of the training loss itself. A corrected equation and released code would resolve the ambiguity, and the empirical evidence is otherwise strong enough that rejection is not warranted; hence the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":21994,"tokens_out":9632,"duration_ms":127211,"concrete_test":"Request the official code and run a controlled comparison on SYSU-MM01 All-Search: train RoDE with Lra exactly as in Eq. (7) and with gamma_i computed from Eq. (18) as written; if the model fails to learn or Rank-1 is far below 62.88, the printed loss is not the evaluated loss. Then train with the corrected generalized CE (1-p^gamma)/gamma and with gamma_i computed from an unweighted per-sample CE loss; if this recovers the reported numbers, the paper's equations must be amended. This one A/B test distinguishes a typographical error from a genuine algorithmic property.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on RAL, the component that is supposed to mitigate noise overfitting. As printed, Eq. (7) defines Lra = -sum p^{gamma_i}(...), with gamma_i in (0,1]. Property 1 then asserts that as gamma_i -> 0 the loss gradually behaves like CE. This is false: for any fixed p in (0,1), p^gamma -> 1 as gamma -> 0, so each summand tends to -1 and the gradient with respect to the logit, -gamma p^gamma (1-p), tends to 0. CE's gradient is p-1, which is nonzero. To obtain the stated limit one needs the normalized generalized CE (1-p^gamma)/gamma, which is not what Eq. (7) writes. Separately, Eq. (9) defines gamma_i from a GMM over per-sample losses ell_i^P, but the text says ell_i^P is calculated by L_i^P in Eq. (18), and Eq. (18) itself contains gamma_i. The weighting is therefore self-referential unless one specifies an unweighted loss for the GMM or a previous-iteration gamma_i, neither of which is stated. Because RAL is retained in every ablation and is the mechanism claimed to handle pseudo-label noise, this internal inconsistency means the published equations do not uniquely determine the training objective that produced the reported SOTA results. This is an internally inconsistent mathematical specification, not merely a disagreement with prior work, and it is more foundational than the acknowledged residual CCM mismatch.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces RoDE, an unsupervised visible-infrared person re-identification framework designed to train under pseudo-label noise. It decomposes the problem into three challenges: noise overfitting, error accumulation, and noisy cluster correspondence. The proposed method combines a Robust Adaptive Learning (RAL) mechanism with a robust loss and adaptive per-sample weights, a Robust Duality Learning (RDL) pipeline that alternately trains two models using each other's pseudo-labels, and a Cluster Consistency Matching (CCM) mechanism that aligns clusters across modalities and models. Experiments on SYSU-MM01, RegDB, and LLCM report state-of-the-art unsupervised results, together with ablations, parameter analysis, and robustness studies. The central claim is that the combination of RAL, RDL, and CCM enables robust learning from pseudo-label noise in UVI-ReID.","tokens_in":22360,"tokens_out":4686,"duration_ms":54094,"significance":"If the learning objective is correctly specified, the paper addresses a relevant and under-explored problem and provides a fairly complete empirical study: three benchmarks, comparisons with 26 methods, ablations of all three components, and dedicated robustness analyses. The dual-model training and cluster-consistency matching are sensible ideas, and the reported gains over prior unsupervised methods are substantial. However, the significance is currently tempered by an internal inconsistency in the core RAL loss: as printed, Eq. (7) does not have the claimed cross-entropy limit, and the adaptive weighting in Eq. (9) is self-referential through Eq. (18). Because these issues affect the mechanism that is central to the paper's claims, the manuscript requires revision before the results can be accepted.","major_comments":[{"comment":"The stated Property 1(1) is mathematically incorrect. For fixed p in (0,1), p^gamma tends to 1 as gamma tends to 0, so each summand of Lra tends to -1 and the gradient with respect to the logit, -gamma p^gamma (1-p), tends to 0; cross-entropy has gradient p-1. To obtain the claimed limit one needs a normalized generalized cross-entropy such as (1-p^gamma)/gamma. Since RAL is the component claimed to mitigate noise overfitting and is retained in every ablation, the published equations do not describe the objective that produced the reported results; please correct Eq. (7) or Property 1 and re-state the actual loss used.","section":"Section III-B, Eq. (7) and Property 1"},{"comment":"The adaptive weight gamma_i is defined via a GMM over per-sample losses ell_i^P, where the text states ell_i^P is calculated by L_i^P in Eq. (18). However, Eq. (18) itself contains p^{gamma_i} with the same gamma_i, making the definition circular unless gamma_i is computed from an unweighted loss or from a previous iteration, which is not stated. Please specify the exact computation order and the loss used for GMM fitting.","section":"Section III-B/E, Eqs. (8)-(9) and Eq. (18)"},{"comment":"The robustness analysis reports that cross-modal and cross-model mismatching rates persist at 15% to 35% even after CCM. Since CCM is what enables the dual-model pipeline, this residual mismatch is substantial, and the paper should provide additional evidence that RAL and RDL tolerate it, for example an oracle-alignment upper bound, a sensitivity analysis with injected mismatch, or an analysis of which samples are reweighted. As written, the claim that noisy cluster correspondence is resolved is stronger than the evidence.","section":"Section IV-H.3, Figure 9(e/f)"}],"minor_comments":[{"comment":"The notation eyP_i and eyQ_i is used for pseudo-labels, but the sentence 'xP_i belongs to the clustering center eyP_i' is imprecise; please define these as pseudo-label indices assigned by clustering.","section":"Section III-A, Eq. (4)"},{"comment":"The text says 'detailed proofs for Property 1 are available in the Appendix,' but no appendix is included in the manuscript; please add the appendix or remove the reference.","section":"Section III-B"},{"comment":"The labels 'RoDE w/o RAL (Lalpha_ra)' and 'RoDE w/o RAL (Lbeta_ra)' are confusing because Lalpha_ra and Lbeta_ra are themselves robust losses; please rename to clarify which RAL component is removed.","section":"Section IV-F, Table III"},{"comment":"The gray shaded 'recommended parameter range' is not quantified; please state the range and the fixed values of other hyperparameters used in the parameter analysis.","section":"Section IV-E, Figure 5"},{"comment":"The loss-distribution plots in Figures 9(a)-(d) would be more informative with quantitative overlap measures, such as the AUC of clean/noisy separation or estimated noise rates, rather than visual inspection alone.","section":"Section IV-H.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of TIFS. The recommended revision is major because the central loss equation needs mathematical correction; once corrected and the adaptive weighting is unambiguously specified, the empirical study could support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is what you should know: the paper reports state-of-the-art numbers for unsupervised visible-infrared person re-identification on SYSU-MM01, RegDB, and LLCM, and the gains look real. But the central robust loss, Eq. (7), is described with a property that its formula does not satisfy, and the adaptive weight gamma_i is defined self-referentially. This makes the method, as written, not a well-defined training objective.\n\nWhat is genuinely new: the paper frames pseudo-label noise as three separate problems — noise overfitting, error accumulation, and noisy cluster correspondence — and builds a system that attacks all three. The dual-model co-training with mutual pseudo-labels is a sensible idea to prevent self-confirmation, and the Cluster Consistency Matching mechanism, which aligns clusters across both modalities and models, is the most interesting component; the ablation shows removing cross-model CCM drops Rank-1 from 62.9 to 3.6 on SYSU-MM01, so that part is doing real work. The empirical study is broad: three datasets, ablations, parameter analysis, t-SNE, and a robustness analysis that honestly reports a persistent 15-35% cluster mismatch.\n\nWhere it falls apart: Property 1 says as gamma_i approaches 0, Lra behaves like cross-entropy. With Lra = -p^gamma, as gamma->0 each term tends to -1, not -log p, and the gradient w.r.t. the logit tends to 0. The version that actually has the CE limit is (1-p^gamma)/gamma. Either the equation is a typo or the property is fabricated; either way, the paper as printed does not specify the loss that produced the results. Second, gamma_i is computed from a GMM on per-sample losses, but the text says those losses are L_i^P from Eq. (18), and Eq. (18) contains gamma_i itself. That is a circular definition unless the authors intended to use an unweighted loss or the previous iteration's weights, neither of which is stated. Since RAL is the core mechanism claimed to handle noise, this is a load-bearing flaw.\n\nRecommendation: send to peer review — the empirical work is solid enough to deserve referee time, and the math issues are likely fixable. But the authors should be required to correct the loss definition, clarify the gamma_i computation, and release code/data so the numbers can be checked. I would not cite this version as a usable method until that happens.","headline":"Solid empirical gains for unsupervised VI-ReID, but the core robust loss is mathematically mis-specified and the method is not reproducible from the written equations.","tokens_in":22881,"tokens_out":5336,"would_cite":false,"duration_ms":60834,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Unsupervised visible-infrared person re-identification can be made robust to pseudo-label noise by combining adaptive sample reweighting, two mutually supervising networks, and cluster alignment; the paper's RoDE framework reports…","keywords":["unsupervised visible-infrared person re-identification","pseudo-label noise","noise overfitting","error accumulation","noisy cluster correspondence","dual-model learning","cluster consistency matching","robust adaptive learning"],"falsifier":"A controlled test would replace CCM's aligned assignments with random permutations of cluster labels across the two models while keeping RAL and the dual-network training intact; if Rank-1 on SYSU-MM01 does not collapse far below the reported 62.88%, then the paper's explanation that cluster alignment is the load-bearing component would be refuted, and the gains would have to be attributed to the reweighting and duality alone.","tokens_in":21815,"feed_emoji":"🔍","tokens_out":8495,"duration_ms":88882,"temperature":0.7,"pith_summary":"Unsupervised visible-infrared person re-identification asks a model to match the same person across daylight and infrared images without any identity labels, so it must invent its own supervision by clustering features. Those invented labels are frequently wrong, and this paper argues that the wrongness is not a minor nuisance but a structured failure with three parts: the model overfits to mislabeled samples, it reinforces its own mistakes over time, and the clusters it builds disagree across modalities and across the networks doing the labeling. The paper's central claim is that a single framework, RoDE, can address all three at once and thereby push unsupervised accuracy well beyond previous methods. RoDE reweights samples by estimated cleanliness, trains two different networks that supervise each other, and aligns their cluster assignments before exchanging labels. On three benchmarks the paper reports state-of-the-art results, including 62.88% Rank-1 and 57.91% mAP on SYSU-MM01 in All-Search mode.","feed_headline":"Dual networks tame noisy labels in cross-modal person matching","feed_subtitle":"Adaptive reweighting and cluster alignment outperform prior unsupervised methods on three benchmarks.","key_machinery":"The load-bearing mechanism is Cluster Consistency Matching (CCM), an alignment step that builds a cost matrix $S_{ij} = \\exp\\left(1 - \\cos(c_i, c_j)\\right)$ between normalized cluster centers and solves a binary linear-sum assignment, applied twice: once to match visible and infrared clusters within each model, and once to match clusters between the two models. This produces the starred pseudo-labels that each network uses to supervise the other. Around it sit two supporting mechanisms: the Robust Adaptive Learning loss $L_{\\mathrm{ra}} = -\\sum_i p_i^{\\gamma_i}$, whose per-sample exponent is set by a two-component Gaussian mixture over losses, and the dual-network alternating training schedule that prevents a single model from accumulating its own errors. The alignment is what converts two independent noisy labelers into a single consistent teacher.","core_discovery":"The paper's contribution is the explicit treatment of Pseudo-Label Noise (PLN) as a three-headed problem in unsupervised visible-infrared re-identification. Its Robust Adaptive Learning (RAL) mechanism uses a two-component Gaussian mixture over per-sample losses to label each sample clean or noisy, then sets a per-sample exponent $\\gamma_i$ in the robust loss $L_{\\mathrm{ra}} = -\\sum p^{\\gamma_i}$ so clean samples are optimized strongly and suspected noisy ones are down-weighted. Its Robust Duality Learning (RDL) trains two networks with different initializations, each optimized on the pseudo-labels produced by the other, so that no single model grows overconfident in its own errors. Its Cluster Consistency Matching (CCM) then aligns the two networks' clusters both across visible and infrared modalities and across the two models, using linear-sum assignment on normalized cosine similarities between cluster centers, before the cross-supervision is applied. The paper shows that removing the cross-model alignment collapses Rank-1 on SYSU-MM01 from 62.88% to 3.56%, which is its evidence that consistent cluster correspondence is what makes the duality strategy viable.","pith_inferences":["Because the residual cluster mismatch rate stays at 15-35%, the obvious next step the authors do not state is to replace the fixed linear-sum assignment with a learned matching cost or a soft alignment, which could raise the ceiling further without changing the dual-training design.","The three-way decomposition of pseudo-label noise—overfitting, error accumulation, and cluster correspondence—is a transferable vocabulary for other unsupervised cross-modal retrieval tasks such as text-image or sketch-photo matching, where cluster pseudo-labels are equally unreliable.","The GMM clean/noisy split assumes per-sample loss correlates with label correctness; if hard but correctly labeled samples consistently receive high loss, the $\\gamma_i$ weighting might suppress useful examples, a boundary that future work could test by inspecting which samples the GMM flags."],"forward_implications":["On SYSU-MM01 All-Search, RoDE reports 62.88% Rank-1 and 57.91% mAP, the best unsupervised result in the paper's comparison and close to several supervised methods.","The ablation in which cross-model CCM is removed drops SYSU-MM01 Rank-1 to 3.56%, implying that consistent alignment between the two supervising networks is necessary for the dual-training strategy to work at all.","On RegDB V2T, RoDE reports 88.77% Rank-1, a 3.18-point improvement over the strongest unsupervised baseline in the table.","The paper's loss distribution analysis shows that dual models separate clean from noisy samples far better than a single model, supporting the claim that error accumulation is mitigated by cross-supervision.","The robustness analysis shows cluster mismatching persists at 15-35%, so the method is tolerant of, rather than immune to, residual alignment errors."],"supporting_citations":[{"why":"Progressive graph matching and alternate learning method whose cluster-alignment strategy RoDE builds on and compares against.","marker":"[2]"},{"why":"GUR, the previous state-of-the-art unsupervised baseline on LLCM that RoDE's reported results surpass.","marker":"[4]"},{"why":"DOTLA, a collaborative pseudo-label refinement method used as a key comparison and in the loss-distribution analysis.","marker":"[6]"},{"why":"Twin noisy labels framework for supervised VI-ReID, which motivates RoDE's joint treatment of label and correspondence noise.","marker":"[8]"},{"why":"IMSL, used to illustrate error accumulation and listed as a comparison method.","marker":"[9]"},{"why":"SCA-RCP, the recent noisy-label-aware unsupervised method that serves as the strongest RegDB baseline.","marker":"[10]"},{"why":"DBSCAN, the clustering algorithm that produces the pseudo-labels and cluster centers RoDE aligns.","marker":"[30]"},{"why":"AGW, the feature extractor backbone used for both visible and infrared modalities.","marker":"[31]"},{"why":"SYSU-MM01, the primary benchmark where RoDE reports 62.88% Rank-1 and 57.91% mAP.","marker":"[45]"},{"why":"RegDB and LLCM, the two additional benchmarks used in the comparison tables.","marker":"[46]"}],"fun_headline_variants":["Dual networks defeat noisy pseudo-labels in cross-modal re-ID","Robust duality learning tames noise in infrared-visible matching","Cluster alignment plus dual models for noise-robust person re-ID","Two networks, cluster matching: no clean labels needed","Pseudo-label noise? Dual learning and cluster matching answer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Cluster Consistency Matching assumes that after normalization a cluster center's nearest counterpart in the other modality or model is usually its true partner, so the assignment step is mostly correct; the paper's own analysis finds a persistent mismatching rate of 15% to 35%.","fun_headline_variants_meta":{"raw":{"variants":["Dual networks defeat noisy pseudo-labels in cross-modal re-ID","Robust duality learning tames noise in infrared-visible matching","Cluster alignment plus dual models for noise-robust person re-ID","Two networks, cluster matching: no clean labels needed","Pseudo-label noise? Dual learning and cluster matching answer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000776,"raw_usage":{"total_tokens":3473,"prompt_tokens":1028,"completion_tokens":2445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":2360}},"tokens_in":644,"tokens_out":2445,"duration_ms":20400,"temperature":1.0,"reasoning_tokens":2360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:48:08.917245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled test would replace CCM's aligned assignments with random permutations of cluster labels across the two models while keeping RAL and the dual-network training intact; if Rank-1 on SYSU-MM01 does not collapse far below the reported 62.88%, then the paper's explanation that cluster alignment is the load-bearing component would be refuted, and the gains would have to be attributed to the reweighting and duality alone.","supporting_citations":[{"cited_title":"Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning,","cited_arxiv_id":null,"evidence_quote":"Progressive graph matching and alternate learning method whose cluster-alignment strategy RoDE builds on and compares against."},{"cited_title":"Towards grand unified representation learning for unsupervised visible-infrared person re-identification,","cited_arxiv_id":null,"evidence_quote":"GUR, the previous state-of-the-art unsupervised baseline on LLCM that RoDE's reported results surpass."},{"cited_title":"Unsupervised visible-infrared person reid by collaborative learning with neighbor- guided label refinement,","cited_arxiv_id":null,"evidence_quote":"DOTLA, a collaborative pseudo-label refinement method used as a key comparison and in the loss-distribution analysis."},{"cited_title":"Learning with twin noisy labels for visible-infrared person re-identification,","cited_arxiv_id":null,"evidence_quote":"Twin noisy labels framework for supervised VI-ReID, which motivates RoDE's joint treatment of label and correspondence noise."},{"cited_title":"Inter-modality simi- larity learning for unsupervised multi-modality person re-identification,","cited_arxiv_id":null,"evidence_quote":"IMSL, used to illustrate error accumulation and listed as a comparison method."},{"cited_title":"Inter-intra modality knowledge learning and clustering noise alleviation for unsupervised visible-infrared person re-identification,","cited_arxiv_id":null,"evidence_quote":"SCA-RCP, the recent noisy-label-aware unsupervised method that serves as the strongest RegDB baseline."},{"cited_title":"A density-based algorithm for discovering clusters in large spatial databases with noise,","cited_arxiv_id":null,"evidence_quote":"DBSCAN, the clustering algorithm that produces the pseudo-labels and cluster centers RoDE aligns."},{"cited_title":"Deep learning for person re-identification: A survey and outlook,","cited_arxiv_id":null,"evidence_quote":"AGW, the feature extractor backbone used for both visible and infrared modalities."},{"cited_title":"Rgb-infrared cross-modality person re-identification,","cited_arxiv_id":null,"evidence_quote":"SYSU-MM01, the primary benchmark where RoDE reports 62.88% Rank-1 and 57.91% mAP."}],"review_version":1}