{"id":"d3d277a7-115c-4a6a-b0b9-4681ef383728","arxiv_id":"2412.12220","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Neighbor-derived soft labels and sample weighting reduce pseudo-label noise in unsupervised visible-infrared person re-identification, improving state of the art on SYSU-MM01 and RegDB.","lead":"This paper presents a method that lets unsupervised visible-infrared person re-identification systems clean up the noisy labels they generate for themselves, by borrowing labels from nearby images. It matters because it improves accuracy on standard night-vision person matching benchmarks with a simple neighbor-based recipe.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed SOTA advantage is not yet established: Table 1's PGM baseline is over 8 points weaker than the authors' own reimplementation in Table 2, and all introduced hyperparameters are tuned on the test sets.","rationale":"The reader's verdict is already CONDITIONAL; our concern reinforces the condition rather than changing it. We focus on the empirical-evidence chain because the paper's headline is an empirical SOTA claim. The Table 1/2 baseline discrepancy is verifiable from the manuscript itself and indicates that the comparison may not be apples-to-apples. Test-set hyperparameter tuning is visible in the supplementary figures and is a known source of inflated benchmark numbers. A single re-run with validation-based tuning and a self-consistent baseline would settle whether the modules genuinely outperform prior work. We do not see an internal algorithmic contradiction; the neighbor assumption is risky, but the ablation results suggest it holds in practice, so we do not elevate it to the primary concern.","tokens_in":13851,"tokens_out":8961,"duration_ms":82727,"concrete_test":"Re-run the full Table 1 comparison under a fixed protocol: use the authors' own reimplementation of PGM (the Table 2 baseline configuration) as the PGM row, and select k, mu, lambda, w on a held-out validation split (e.g., 20% of training identities) instead of on the test set. If the final model no longer exceeds the strongest baseline/SOTA under this protocol, the central claim fails; if it still wins by comparable margins, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that N-ULC and N-DW deliver state-of-the-art results on SYSU-MM01 and RegDB. That claim is supported by Table 1, but two protocol problems compromise its reliability. First, the PGM baseline row in Table 1 (RegDB V-to-I: 69.48 r1; IR-to-V: 69.85 r1) is markedly weaker than the same PGM baseline used in the ablation study (Table 2 row 1: 77.92 and 77.44 r1 on those settings). The comparison table therefore does not use the authors' own baseline, making the 'outperforms existing approaches' conclusion dependent on published numbers that may have been produced under different settings. Second, all four introduced hyperparameters (k, mu, lambda, w) are tuned directly on the test sets (Figs. 3 and 5, Table 3), so the reported numbers are test-set-selected best results, not an unbiased evaluation. Together these issues mean the size and even the existence of the claimed advantage over prior work is not yet established under a consistent, unbiased protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for unsupervised visible-infrared person re-identification (USL-VI-ReID) that mitigates pseudo-label noise using neighbor information. Two modules are introduced: Neighbor-guided Universal Label Calibration (N-ULC), which replaces hard pseudo labels with soft labels derived from the Jaccard overlap between k-nearest neighbors and clusters in both intra- and inter-modality spaces, and Neighbor-guided Dynamic Weighting (N-DW), which down-weights unreliable samples based on neighbor consistency. The method is built on the PGM baseline with an alternating learning scheme. Experiments on RegDB and SYSU-MM01 report state-of-the-art results, e.g., 88.75 r1 on RegDB V-to-I and 61.81 r1 on SYSU-MM01 all-search. The paper also provides a Rademacher generalization bound for soft-label learning, citing Wei et al. 2022.","tokens_in":14088,"tokens_out":5392,"duration_ms":45171,"significance":"The proposal is simple and the ablation in Table 2 indicates each module contributes to improved performance over the authors' PGM reimplementation. The code is released, which is a strength. If the improvements are reliable, the work provides a practical way to reduce label noise in unsupervised cross-modality re-ID. However, the empirical evidence faces two protocol issues that affect the validity of the claimed SOTA: the comparison table uses a PGM baseline different from the ablation baseline, and hyperparameters are tuned on test sets. These issues need to be addressed.","major_comments":[{"comment":"The PGM baseline in the comparison table is inconsistent with the PGM baseline used in the ablation study. In Table 1, PGM (CVPR-23) reports 69.85 r1 (IR-to-V) and 69.48 r1 (V-to-I) on RegDB, while Table 2's first row (PGM reimplementation) reports 77.44 r1 and 77.92 r1 for the same settings. The reported improvement of Ours over PGM in Table 1 is therefore substantially larger than the actual gain over the authors' own reimplementation (88.75 vs 77.92 on V-to-I). The authors should reconcile this discrepancy, e.g., by reporting their reimplemented PGM in Table 1 (and possibly other baselines) or by explaining the setting differences, to make the SOTA claim apples-to-apples.","section":"Table 1 and Table 2"},{"comment":"The hyperparameters k, mu, lambda, and w are selected based on performance on the test sets (e.g., k is chosen on SYSU-MM01 all-search and RegDB V-to-I; mu, lambda, and w on SYSU-MM01 all-search). This constitutes test-set overfitting and the reported results are the best over the searched values, not unbiased estimates. The authors should use a held-out validation split for hyperparameter selection, or report the variance across values and justify the choice without reference to test performance.","section":"Figures 3, 5 and Table 3"}],"minor_comments":[{"comment":"The notation P_intra and P_inter is used for both the raw Jaccard values and the ℓ1-normalized distribution; please clarify the normalization step explicitly.","section":"Eqs. (5), (8)"},{"comment":"The footnote 'GUR† deontes results without camera information' contains a typo: 'denotes' is misspelled as 'deontes'.","section":"Table 1 caption"},{"comment":"The module name 'Neighbor-Guided Label Universal Calibration' in the caption is inconsistent with the section title 'Neighbor-Guided Universal Label Calibration'; use a consistent name throughout.","section":"Table 2 caption"},{"comment":"The Discussion paragraph claims a curriculum-learning effect ('we assign higher weights to easily labeled samples... more challenging samples are assigned greater weights if they demonstrate increased consistency'), but Eq. (11) computes weights as a static function of current neighbor statistics; the claimed curriculum mechanism is not explicitly implemented or tested.","section":"Section 'Neighbor-Guided Dynamic Weighting'"},{"comment":"Two Yin et al. 2024 entries (2024a and 2024b) are from different research groups; this is confusing in the citation list and should be disambiguated by author names or titles.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The baseline discrepancy between Table 1 and Table 2 is the most serious issue. If the authors can provide a consistent evaluation protocol—ideally reimplementing all baselines under the same settings, and using a validation split for hyperparameter selection—the paper would be publishable. The method itself is simple and the ablation is encouraging. The Rademacher section is superficial and should either be removed or made relevant. The self-citations by the corresponding author in the label-noise related work are not problematic by themselves, but the authors should ensure the novel contribution is clearly differentiated from prior neighbor-based USL-VI-ReID methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes two neighbor-based modules on top of the PGM baseline: N-ULC replaces hard pseudo labels with soft labels derived from neighbor overlap with clusters, in both homogeneous (within-modality) and heterogeneous (cross-modality) learning; N-DW weights samples by neighbor consistency. The idea is simple and the ablation is the strongest part: adding the modules to their own PGM reimplementation lifts RegDB V-to-I rank-1 from 77.92 to 88.75, and SYSU-MM01 all-search from 54.14 to 61.81. That is a solid, consistent gain, and the modules are clearly described.\n\nWhat's genuinely new: neighbor-guided label refinement exists in DOTLA and in arXiv:2405.05613, but those apply it only to cross-modality correspondences. Applying the same neighbor-derived soft labels uniformly to both intra- and inter-modality prototype learning, plus the weighting scheme, does appear to be a new combination. The paper gives credit to those related works.\n\nThe soft spots are real but not fatal, and they are concentrated in the empirical protocol. The main comparison table reports a PGM baseline of 69.48/69.85 rank-1 on RegDB, while the ablation's own baseline is 77.92/77.44. That is an 8-point gap, and it means the claim \"outperforms existing approaches\" is not apples-to-apples unless the PGM numbers come from identical settings. The stress-test note is right on this. Also, the hyperparameters k, mu, lambda, w are tuned on the test sets via Fig. 3/5 and Table 3. That inflates the reported figures and makes them test-set-selected best results rather than unbiased estimates. Neither flaw is unusual in this literature, but combined they undermine the SOTA claim.\n\nThe theoretical analysis section is a borrowed Rademacher bound for label smoothing (Wei et al. 2022) with no adaptation to neighbor-derived soft labels or to the cross-modal setting. It does not hurt the method but adds little.\n\nThe implementation seems straightforward and the code is promised; I didn't verify the repo. Overall, the paper is worth a serious referee. The likely outcome should be a revise: fix the baseline consistency, report hyperparameter selection on a validation split or at least acknowledge the limitation, and trim or rework the theory. The core method is likely sound and the community would benefit from an honest comparison.\n\nFor a reading group, it's a useful example of a simple pseudo-label denoising recipe; I'd bring it in if anyone works on unsupervised cross-modal learning.\n\nRecommendation: send to peer review with a request for a careful comparison protocol. The paper is not ready as-is, but it deserves the referee time.","headline":"Neighbor-based soft labeling for USL-VI-ReID shows a clean ablation win, but the headline SOTA claim rests on a mismatched baseline and test-set-tuned hyperparameters.","tokens_in":14673,"tokens_out":2178,"would_cite":true,"duration_ms":18782,"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":"This paper claims that in unsupervised visible-infrared person re-identification, replacing hard cluster pseudo-labels with soft labels inferred from each sample's k nearest neighbours—both within a modality and across modalities—and…","keywords":["unsupervised visible-infrared person re-identification","pseudo-label noise","neighbor-guided label calibration","soft pseudo labels","sample weighting","progressive graph matching","cross-modality learning","visible-infrared ReID"],"falsifier":"On a training set whose ground-truth identities are known, measure the 'neighbour purity' of each sample—the fraction of its $k$ nearest neighbours that share its true identity—throughout training in both modalities. If, for a substantial fraction of samples, the majority identity of the neighbour set disagrees with the cluster pseudo-label (or with the true identity) at the point where N-ULC and N-DW are active, then the soft labels and weights would be amplifying the very noise they are designed to remove. A sharper test: replace the feature-space neighbour sets with random neighbour sets and check that the reported gains vanish, or replace them with oracle neighbour sets defined by true identities and check that the gains sharply increase.","tokens_in":13636,"feed_emoji":"👤","tokens_out":11795,"duration_ms":91925,"temperature":0.7,"pith_summary":"Unsupervised visible-infrared person re-identification must match people across daylight and thermal cameras without identity annotations, so methods cluster features and treat each cluster as a pseudo-identity. Those pseudo-labels are noisy: the same person's images frequently land in different clusters within one modality, and cluster correspondences across modalities are noisier still. The paper argues that this universal label noise can be relieved by asking each sample's $k$ nearest neighbours: the overlap between a sample's neighbour set and each cluster yields a soft label, and the consistency of neighbours' cluster memberships tells the algorithm how much to trust the sample. Added to the Progressive Graph Matching baseline, the proposed N-ULC and N-DW modules raise RegDB visible-to-infrared rank-1 to 88.75% and mAP to 82.14%, and SYSU-MM01 all-search rank-1 to 61.81% and mAP to 58.92%, outperforming existing unsupervised visible-infrared re-identification methods.","feed_headline":"Neighbor soft labels lift unsupervised VI-ReID to 88.75 top-1","feed_subtitle":"By inferring identities from neighbors, it outperforms prior unsupervised methods on RegDB and SYSU-MM01.","key_machinery":"The two load-bearing objects are the neighbour-guided calibrated soft label and the neighbour-consistency weight. For a query $q$ with feature set $U$, the paper takes its $k$ nearest neighbours $N(q, U, k)$ and measures the correlation of $q$ with each cluster $C_l$ by the normalized overlap $|N(q, U, k) \\cap C_l| \\,/\\, |N(q, U, k) \\cup C_l|$. After $\\ell^1$-normalizing this correlation vector into $P_q$, the hard one-hot pseudo-label $I_q$ is blended into the soft label $\\tilde{I}_q = \\mu I_q + (1-\\mu)P_q$, which replaces the hard label in both the homogeneous and heterogeneous contrastive losses. The companion weight $\\omega_q = \\exp(-w(1-[P_q]_{\\text{label}})^2)$ down-weights each sample's contribution when its neighbours disagree with the cluster it was assigned to. The whole mechanism runs on the Progressive Graph Matching baseline, which supplies the clusters, prototypes, and cross-modality cluster correspondences that define the losses being calibrated.","core_discovery":"The central discovery is that universal label noise—noise in pseudo-labels both within a modality (homogeneous) and across modalities (heterogeneous)—can be substantially reduced without extra supervision by deriving identity evidence from neighbours instead of relying on cluster assignments alone. The paper shows that a soft label formed by the normalized overlap between a query's $k$-nearest-neighbour set and every cluster, blended with the original one-hot cluster label as $\\tilde{I} = \\mu I + (1-\\mu)P$, represents the query's true identity more faithfully than the hard label. The same construction transfers to the other modality by collecting neighbours across the visible and infrared feature sets. The paper further shows that the overlap at the query's own cluster provides a per-sample reliability weight $\\omega = \\exp(-w(1-[P]_{\\text{label}})^2)$, and that weighting the soft-label losses by this reliability stabilises training. In the reported experiments, the combination lifts the PGM baseline on RegDB and SYSU-MM01 and surpasses earlier USL-VI-ReID methods.","pith_inferences":["The neighbour-overlap recipe is not tied to the PGM baseline: any clustering-based USL-VI-ReID pipeline that maintains per-modality features and cluster assignments could replace its hard pseudo-labels with N-ULC-style soft labels and add N-DW-style weights, so the modules are likely to transfer to other frameworks.","Weighting neighbours by feature distance, or restricting the neighbour set to reciprocal neighbours (neighbours that also list the query), could produce sharper soft labels than the unweighted Jaccard ratio used here; that variant is untested in the paper.","The method's weakest point is early-training neighbour purity; a diagnostic that tracks the agreement between neighbour-set majorities and cluster pseudo-labels on a small labelled subset could predict when the modules help and might be used to schedule the weighting strength $w$ adaptively.","Because the soft labels are built entirely from the current feature space, the method should benefit from any improvement in the backbone; combining N-ULC and N-DW with stronger pretrained encoders is a natural and untested extension."],"forward_implications":["Ablations show that each module contributes: adding only N-ULC to PGM improves SYSU-MM01 all-search rank-1 from 54.14 to 58.54, adding only N-DW improves it to 55.87, and the combination reaches 61.81.","Because the soft label is a blend rather than a replacement, the model can still exploit the cluster structure while tolerating individual assignment mistakes, so training is less sensitive to the DBSCAN distance thresholds.","The neighbour-consistency weight acts as a curriculum: early in training, samples whose neighbours agree with their cluster dominate the loss, and as features improve, harder samples are gradually trusted.","The same default hyper-parameters ($\\mu=0.7$, $\\lambda=3$, $w=10$, and modest $k$) work across both RegDB and SYSU-MM01, indicating the gains do not come from per-dataset tuning.","Performance on RegDB visible-to-infrared reaches 88.75% rank-1 and 82.14% mAP, numbers above all previously reported USL-VI-ReID results listed in the paper, including several that use extra labelled visible data."],"supporting_citations":[{"why":"Supplies the Progressive Graph Matching baseline; its clusters, prototypes, and cross-modality label transformers define the losses that N-ULC and N-DW calibrate and reweight.","marker":"(Wu and Ye 2023)"},{"why":"DBSCAN is the clustering algorithm the paper uses to generate the pseudo-labels whose noise is the target of the method.","marker":"(Ester et al. 1996)"},{"why":"Theorem 1 of this work supplies the Rademacher generalization bound the paper invokes to justify optimizing with soft probabilistic labels.","marker":"(Wei et al. 2022)"},{"why":"Defines the RegDB benchmark dataset used for the reported experiments.","marker":"(Nguyen et al. 2017)"},{"why":"Defines the SYSU-MM01 benchmark dataset and the cross-modality re-identification task that the paper builds on.","marker":"(Wu et al. 2017)"},{"why":"ADCA is a leading USL-VI-ReID baseline, and the paper follows its two-stage training protocol and compares against its reported results.","marker":"(Yang et al. 2022)"},{"why":"A prior neighbor-guided label-refinement method for USL-VI-ReID that the paper compares with and distinguishes from its own more universal use of neighbors.","marker":"(Cheng et al. 2023b)"},{"why":"Channel Augmentation is used during training as part of the experimental protocol for handling the visible-infrared modality gap.","marker":"(Ye et al. 2021a)"}],"fun_headline_variants":["Neighbor soft labels cut label noise for unsupervised VI-ReID","Boost unsupervised VI-ReID: infer identity from neighbors","Neighbor-guided label calibration beats prior USL-VI-ReID","Neighbor soft labels outperform on RegDB and SYSU-MM01","Unsupervised VI-ReID gets neighbor-powered label fix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the $k$ nearest neighbours of a sample, computed in the current feature space, mostly share that sample's true identity, so the overlap between the neighbour set and the clusters is an accurate proxy for identity confidence; if early features are poor or the clusters are badly mixed, the neighbour sets carry the same noise the method aims to remove.","fun_headline_variants_meta":{"raw":{"variants":["Neighbor soft labels cut label noise for unsupervised VI-ReID","Boost unsupervised VI-ReID: infer identity from neighbors","Neighbor-guided label calibration beats prior USL-VI-ReID","Neighbor soft labels outperform on RegDB and SYSU-MM01","Unsupervised VI-ReID gets neighbor-powered label fix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1325,"prompt_tokens":983,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":257}},"tokens_in":599,"tokens_out":342,"duration_ms":3654,"temperature":1.0,"reasoning_tokens":257,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:56:29.216408+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a training set whose ground-truth identities are known, measure the 'neighbour purity' of each sample—the fraction of its $k$ nearest neighbours that share its true identity—throughout training in both modalities. If, for a substantial fraction of samples, the majority identity of the neighbour set disagrees with the cluster pseudo-label (or with the true identity) at the point where N-ULC and N-DW are active, then the soft labels and weights would be amplifying the very noise they are designed to remove. A sharper test: replace the feature-space neighbour sets with random neighbour sets and check that the reported gains vanish, or replace them with oracle neighbour sets defined by true identities and check that the gains sharply increase.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Theorem 1 of this work supplies the Rademacher generalization bound the paper invokes to justify optimizing with soft probabilistic labels."},{"cited_title":"T.; Hong, H","cited_arxiv_id":null,"evidence_quote":"Defines the RegDB benchmark dataset used for the reported experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ADCA is a leading USL-VI-ReID baseline, and the paper follows its two-stage training protocol and compares against its reported results."}],"review_version":1}