{"id":"dd77b0de-b97a-417d-9a2a-7679b417cbdb","arxiv_id":"2504.14877","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CoEN boosts multi-spectral vehicle ReID by generating a fused proxy from all spectra, ranking each spectrum by similarity to that proxy to select the primary, and enhancing all spectra with two cross-attention strategies.","lead":"This paper presents CoEN, a network that improves multi-spectral vehicle re-identification when some camera spectra (visible, near-infrared, thermal) are degraded. It fuses all spectra into a proxy feature, uses proxy-to-spectrum similarity to pick the best-quality spectrum, and enhances every spectrum with cross-attention, achieving new state-of-the-art numbers on three benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DQSM ranks spectra by cosine similarity to a proxy built from the very same spectra (Eqs. 2-6), so the dynamic-selection mechanism is unvalidated and may reduce to fixed TIR selection; the claimed 0.8-1.0 mAP dynamic advantage is within single-run noise.","rationale":"The reader's weakest assumption matches my own: DQSM's quality ranking is self-referential because the proxy is produced from the same spectral features being scored. This is load-bearing because the paper's claimed contribution is dynamic selection of the primary spectrum; if the ranking is uninformative, the primary-based enhancement branch is just noise injection and the state-of-the-art claim rests only on proxy fusion plus extra capacity. The paper's ablations are internally plausible but do not isolate the ranking: Table 6 compares dynamic selection against fixed RGB/NIR/TIR primaries, but the margin over fixed TIR is 0.8 mAP with no error bars, and the comparison against w/o DQSM is 1.0 mAP. No analysis of which spectrum DQSM actually selects, nor any validation against ground-truth quality, is reported. The concern is not that the method is wrong but that the central mechanism is unsupported; the remedy is a concrete validation experiment plus, ideally, variance reporting. I do not see a stronger objection: the architecture is coherent, the loss functions are standard, and the benchmark comparisons are plausible conditional on reproducible code. The reader's CONDITIONAL verdict remains appropriate, so no change is needed.","tokens_in":19350,"tokens_out":3483,"duration_ms":34368,"concrete_test":"On WMVeID863, log DQSM's argmax selection for every query and compare it with per-query ground-truth quality labels, e.g., the spectrum whose isolated branch yields the highest retrieval score for that query or an image-quality/SNR estimate, and report the selection distribution by scene category (normal, flare, low-light). If the selected primary is almost always TIR, or if agreement with the ground-truth best spectrum is near chance, the dynamic-selection claim is not supported; if selection varies meaningfully with scene and agrees with the oracle, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is DQSM's 'accurate' dynamic selection of the primary spectrum, but the quality score in Eqs. (4)-(6) is cosine similarity between each spectrum's embedding tokens F_R/F_N/F_T and the proxy F_P, and F_P is generated by Eq. (3) from exactly those same embeddings (F_P = ProjB(ProjA(Concat(F_R,F_N,F_T)))). The score is therefore not an independent quality measurement: a spectrum can rank high merely because the progressive projection preserves its own content or because of per-spectrum norm/scale differences, not because it retains more identity cues. The paper provides no ground-truth quality labels, no report of how often DQSM selects each spectrum, and no per-scene breakdown (normal/flare/low-light), even though those scenes are the motivation. Table 6 shows the entire dynamic-selection benefit: w/ DQSM 71.4 mAP vs fixed TIR 70.6 and w/o DQSM 70.4, a 0.8-1.0 point gap with no error bars or repeats. If DQSM's ranking is uninformative, CEM's primary-based enhancement branch (Eqs. 8-9) injects noise into F'_2nd and F'_3rd, and the claimed reason for CoEN's gains collapses to the proxy branch plus more parameters. This does not refute the headline numbers but makes the mechanism claim unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Collaborative Enhancement Network (CoEN) for multi-spectral vehicle re-identification under low-quality, scene-varying degradations. A shared Vision Transformer extracts per-spectrum embedding tokens; the Proxy Generator (PG) concatenates these tokens and applies two progressive projections to form a fused proxy feature. The Dynamic Quality Sort Module (DQSM) ranks RGB, NIR, and TIR spectra by the mean cosine similarity between each spectrum's embedding tokens and the proxy, and selects the highest-scoring spectrum as the primary spectrum. The Collaborative Enhancement Module (CEM) then enhances the remaining spectra via two cross-attention branches, one using the primary spectrum as key/value and one using the proxy, with residual aggregation. Training uses identity and triplet losses on the backbone, CEM, and proxy branches. The paper reports state-of-the-art or competitive results on WMVeID863, RGBNT100, and MSVR310, together with ablations, hyperparameter analysis, and qualitative visualizations.","tokens_in":19559,"tokens_out":8422,"duration_ms":82601,"significance":"If the reported results hold, CoEN would be a practically useful state-of-the-art method on three public multi-spectral vehicle ReID benchmarks, and its central idea of making primary-spectrum selection scene-adaptive is a reasonable and timely research direction. The paper's strengths are its external comparisons to previously published methods, its internally consistent ablations in which each added component improves mAP, and its clear architectural description. The significance is currently conditional, however, because the paper's claimed novelty rests on DQSM's ability to measure spectral quality, and that measurement is not validated independently of the proxy that DQSM itself builds from the spectra being scored. The headline numerical gains over fixed TIR selection are also reported as single runs without error bars, so the magnitude of the dynamic-selection benefit is not yet established.","major_comments":[{"comment":"The DQSM quality score is not an independent measure of spectral quality. The proxy F_P in Eq. (3) is generated by a learned projection of the concatenation of the very embeddings F_R, F_N, and F_T that Eqs. (5)-(6) then score against, so the correlation score Q_R measures agreement with a feature that is itself a function of F_R. If the projections are approximately linear, F_P is a mixture of the three spectra, and a spectrum can rank first because its contribution to the mixture or its per-token norm is largest, not because it retains more identity cues. The paper provides no ground-truth quality labels, no per-scene selection statistics, and no report of how often each spectrum is chosen as primary; Table 6 shows only a 0.8-1.0 mAP gap over fixed TIR selection with no error bars. The central 'accurate dynamic primary selection' claim is therefore unsupported as stated. A useful control would be to compare DQSM against random or fixed orderings with identical seeds, and to report selection frequencies per scene category; the authors should also clarify whether DQSM is intended as an intrinsic quality estimator or as a consensus-with-proxy measure.","section":"§3.3, Eqs. (2)-(7), Table 6"},{"comment":"The paper reports no error bars or multiple-seed statistics for any of its quantitative claims, and the two hyperparameters gamma (Eq. (9)) and lambda (Eq. (16)) are tuned on WMVeID863 and then evaluated on the same dataset. Table 1's 1.6% mAP advantage over FACENet and Table 6's 0.8% dynamic-selection advantage are of the same order as typical run-to-run variance in ReID experiments, so the current evidence does not exclude the possibility that the main margins are partly noise. I request at least three independent runs with mean and standard deviation for the principal tables, or an explicit evaluation protocol with a held-out validation split for hyperparameter selection. Without this, the ablations in Tables 3, 6, and 7 cannot be reliably used to attribute the gains to DQSM and CEM.","section":"§4.2, §4.5, Tables 1, 3, 6"},{"comment":"The paper does not describe how the discrete sorting operation in Eq. (7) interacts with end-to-end training, nor whether the selected primary spectrum is stable across training epochs. If the ranking changes frequently for a given sample, the enhancement target of the primary-based branch is noisy, and the benefits of dynamic selection could be diluted. At minimum, the authors should include a training-time analysis of selection stability, and in the experiments report how often each spectrum is chosen as primary on WMVeID863 overall and separately for normal, flare, and low-illumination scenes. This would also address the question of whether DQSM is truly necessary or whether a fixed prior such as FACENet's TIR selection is sufficient in practice; the current Table 6 comparison is too coarse to resolve that question.","section":"§3.3, Eq. (7) and §3.4, Eqs. (8)-(11)"}],"minor_comments":[{"comment":"The dataset name is spelled 'MWVeID863' at the start of Section 4.1; it should be 'WMVeID863' to match the rest of the paper.","section":"§4.1, first paragraph"},{"comment":"The subscript '3nd' in Eq. (12) is a typo and should be '3rd'.","section":"Eq. (12)"},{"comment":"The text and figure caption refer to 'Demo' when the cited method is 'DeMo'; please unify the name.","section":"Figure 10 and §4.6"},{"comment":"The dropout notation Dropout_γ is used before it is defined in the text; define gamma in a sentence preceding Eq. (9) or move the definition earlier.","section":"§3.4, Eq. (9)"},{"comment":"The Data Availability Statement says the data are included in the paper, but the experiments use third-party benchmark datasets; the statement should identify the public sources and access information instead.","section":"Data Availability Statement"},{"comment":"The tables would be clearer if the multiple-seed/error-bar columns were added, and if the CLIP variant of CoEN on RGBNT100 were explicitly marked as not exceeding IDEA (87.2 mAP) in the caption, since the current text only describes it as competitive.","section":"Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is the DQSM validity concern: the paper's novelty claim is substantially about dynamic primary selection, but the only evidence is a small single-run mAP gap with no independent quality validation. I would be willing to review a revision that adds selection-frequency statistics, per-scene breakdowns, multi-seed results, and a clear statement of what DQSM is designed to measure. The architecture and external comparisons are plausible, so I do not see this as a rejection; the needed work is substantial enough to require a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental paper for multi-spectral vehicle ReID. The dynamic primary-selection idea (DQSM) is new, and the architecture is sensible. But the evidence that the selection works as claimed is thin: the quality score is self-referential, the gain over a fixed TIR primary is small, there are no error bars, and the code is only promised.\n\nWhat is actually new: DQSM replaces the fixed TIR prior in FACENet with a correlation-based ranking against a fused proxy feature. The two-branch enhancement in CEM, primary-based plus proxy-based cross-attention, is a reasonable design. The paper is clearly written, and the ablations in Tables 3-7 are internally consistent: each component adds a small positive increment, and the comparison set includes CLIP-based baselines. The authors are also honest about not achieving SOTA on RGBNT100 when using a CLIP backbone.\n\nThe main soft spot is the quality score itself. It measures cosine similarity between each spectrum and the proxy, but the proxy is built from those same spectra via a learned projection (Eqs. 2-3). The ranking is therefore not an independent quality measure; it can be driven by projection artifacts or norm differences. The paper never validates the ranking against ground-truth quality, never reports selection frequency by scene type (normal, flare, low-light), and the only evidence is end-to-end mAP. Table 6 shows the dynamic benefit over a fixed TIR primary is 0.8 mAP (71.4 vs 70.6), which is within typical run-to-run noise for a dataset of this size. The complete absence of error bars or multiple-seed runs is a real weakness for a claim of SOTA by 1.6 to 4.9 mAP.\n\nA second issue: gamma and lambda are tuned on WMVeID863 and reported on the same dataset (Sec. 4.5). That is a mild selection-on-test-set problem; they should either fix the values a priori or show sensitivity on the other two benchmarks.\n\nThird, the code is promised at a GitHub URL but not actually released, so reproducibility depends on a future link.\n\nNone of this refutes the headline numbers; the benchmark comparisons are external and the ablations are coherent. But the mechanism story—that DQSM 'accurately selects' the highest-quality spectrum—is unsupported. The paper would be stronger with a per-scene breakdown of which spectrum is selected, a comparison against a learned quality estimator, or simply softer language about what the correlation score means.\n\nThis paper is for researchers working on multi-spectral or low-quality vehicle ReID. It is a niche but active area, and the dynamic-primary idea is a legitimate addition. I would send it to peer review rather than desk reject, but a serious referee should ask for multiple runs, selection validation, and code. If those come back, the paper could be a solid conditional accept.","headline":"A plausible incremental method with a genuinely new dynamic primary-selection idea, but the mechanism is under-validated and the headline gains lack error bars; worth peer review, not a desk reject.","tokens_in":20291,"tokens_out":6186,"would_cite":true,"duration_ms":51800,"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":"CoEN fuses RGB, near-infrared, and thermal spectra into a proxy that selects the best primary spectrum, lifting multi-spectral vehicle re-identification accuracy on three benchmarks.","keywords":["multi-spectral vehicle re-identification","low-quality spectra","proxy feature generation","dynamic quality sort","collaborative enhancement","cross-attention enhancement","Vision Transformer","flare and low-illumination scenes"],"falsifier":"Record the primary spectrum chosen by DQSM for each sample on WMVeID863 and check whether the choice varies across flare, normal, and low-light scenes; if the selected primary is almost always TIR, or if the DQSM ranking does not change when the proxy is computed without the spectrum being scored, then the dynamic quality sort is not driving the reported gains.","tokens_in":18979,"feed_emoji":"🚗","tokens_out":6167,"duration_ms":55327,"temperature":0.7,"pith_summary":"The paper proposes Collaborative Enhancement Network (CoEN) for re-identifying vehicles from RGB, near-infrared, and thermal images when some spectra are degraded by flares, low light, or other conditions. The central claim is that instead of trusting a fixed primary spectrum, the network should build a fused proxy from all spectra, use that proxy to score which spectrum currently carries the most identity information, and then enhance every spectrum using both the best spectrum and the proxy. On the flare-heavy WMVeID863 benchmark, CoEN reports 71.4% mAP and 79.2% Rank-1, exceeding the prior FACENet by 1.6 and 2.2 points respectively, and on RGBNT100 and MSVR310 it reports larger gains over the previous best DeMo. The intended contribution is a scene-adaptive answer to the question of which spectrum should lead the enhancement, rather than a hand-picked prior.","feed_headline":"A fused proxy picks the best spectrum for vehicle re-ID","feed_subtitle":"CoEN selects the primary spectrum per sample and enhances all spectra, beating fixed-spectrum methods on three multi-spectral benchmarks.","key_machinery":"The central object is the proxy feature $F_P = \\mathrm{ProjB}(\\mathrm{ProjA}(\\mathrm{Concat}(F_R^{emb}, F_N^{emb}, F_T^{emb})))$, a fused representation formed by concatenating the embedding tokens of all three spectra and reducing their dimension in two steps. The proxy acts as a shared currency: the Dynamic Quality Sort Module uses token-level cosine similarities between $F_P$ and each spectrum's embedding to assign quality scores and rank the spectra, while the Collaborative Enhancement Module uses $F_P$ as key and value in a proxy-based cross-attention branch. The same proxy is also used as a branch for final identity inference. This design replaces any fixed choice of primary spectrum with a per-sample, proxy-driven selection.","core_discovery":"The paper claims that low-quality multi-spectral vehicle re-identification is best served by a collaboratively enhanced representation, not by enhancing all spectra from one presumed-good spectrum. CoEN builds a proxy feature by progressively fusing embedding tokens from RGB, NIR, and TIR, then ranks spectra by the average cosine similarity of their tokens to the proxy and selects the top-ranked spectrum as primary. The Collaborative Enhancement Module applies cross-attention twice: primary-based enhancement transfers detailed cues from the primary to the two lower-ranked spectra, and proxy-based enhancement supplements every spectrum, including the primary, with fused discriminative cues; residual connections preserve spectrum-specific detail. With a shared Vision Transformer backbone and identity plus triplet losses, the paper reports state-of-the-art results across WMVeID863, RGBNT100, and MSVR310.","pith_inferences":["A natural testable extension is to compare the DQSM's cosine-similarity ranking against ground-truth quality labels or no-reference image-quality scores per spectrum; if the ranking agrees with those labels better than a fixed thermal prior does, the dynamic selection claim is strengthened.","Because the proxy is built from the same spectra it scores, the quality scores may reflect which spectrum the fusion projection preserves rather than true scene-adapted quality; an editor's suggested experiment is to train the proxy on only two spectra and score the held-out third spectrum, checking whether the ranking still tracks known degradations.","The collaborative-enhancement design could plausibly transfer to other multi-modal recognition tasks, such as visible-infrared person re-identification or autonomous driving perception, where the best modality varies with scene conditions and a proxy could serve the same role as a scene-adaptive selector."],"forward_implications":["If the reported gains hold, multi-spectral vehicle re-identification no longer needs a hand-picked primary spectrum such as thermal, because CoEN selects the primary per sample from the proxy ranking.","The two-pronged enhancement strategy (primary-based plus proxy-based cross-attention) is designed to work when all spectra are low quality, a setting where fixed-primary methods degrade, and the paper's ablations support this on the flare-heavy WMVeID863 benchmark.","Because the proxy itself carries more identity information than any single spectrum, the paper shows that adding the proxy as an inference branch improves retrieval even after enhancement, from 69.7% to 71.4% mAP on WMVeID863.","CoEN with a CLIP backbone also improves MSVR310 mAP to 52.2% versus 49.2% for DeMo, suggesting the collaborative enhancement transfers to large pre-trained backbones without requiring text annotations.","Ablation results indicate that progressive projection in the Proxy Generator outperforms element-wise summation or single-step projection, and that dynamic DQSM selection outperforms fixing RGB, NIR, or TIR as the primary spectrum."],"supporting_citations":[{"why":"Introduces the WMVeID863 dataset and FACENet, the fixed-TIR-primary baseline that CoEN reports outperforming by 1.6 mAP and 2.2 Rank-1.","marker":"Zheng et al. (2025)"},{"why":"DeMo is the strongest Transformer-based prior on RGBNT100 and MSVR310, and the paper reports CoEN surpassing it by 2.2 and 4.9 mAP respectively.","marker":"Wang et al. (2025b)"},{"why":"Introduces the RGBNT100 multi-spectral vehicle ReID benchmark and the HAMNet baseline.","marker":"Li et al. (2020)"},{"why":"Introduces the MSVR310 benchmark and the CCNet method, providing the evaluation protocol and a comparison point.","marker":"Zheng et al. (2023b)"},{"why":"The Vision Transformer architecture serves as the shared feature extraction backbone for all three spectra.","marker":"Dosovitskiy et al. (2020)"},{"why":"TOP-ReID is a token-permutation multi-spectral enhancement method used as a comparison and as the TPM alternative in the CEM ablation.","marker":"Wang et al. (2024)"}],"fun_headline_variants":["CoEN fuses spectra to pick the best and boost all for vehicle re-ID","Collaborative proxy guides spectrum choice and enhances all for re-ID","One fused proxy ranks spectra and upgrades low-quality ones for vehicle re-ID","Vehicle re-ID: collaborative enhancement beats fixed-spectrum methods","Dynamic proxy rank picks primary spectrum, enhancing all for re-ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a spectrum's similarity to the merged proxy tells you how much identity information that spectrum still has, even though the proxy was built from the very same spectra; if the similarity merely reflects the way the merging was constructed, the dynamic primary selection loses its advantage.","fun_headline_variants_meta":{"raw":{"variants":["CoEN fuses spectra to pick the best and boost all for vehicle re-ID","Collaborative proxy guides spectrum choice and enhances all for re-ID","One fused proxy ranks spectra and upgrades low-quality ones for vehicle re-ID","Vehicle re-ID: collaborative enhancement beats fixed-spectrum methods","Dynamic proxy rank picks primary spectrum, enhancing all for re-ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3127,"prompt_tokens":987,"completion_tokens":2140,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2049}},"tokens_in":603,"tokens_out":2140,"duration_ms":14811,"temperature":1.0,"reasoning_tokens":2049,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:07.231764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the primary spectrum chosen by DQSM for each sample on WMVeID863 and check whether the choice varies across flare, normal, and low-light scenes; if the selected primary is almost always TIR, or if the DQSM ranking does not change when the proxy is computed without the spectrum being scored, then the dynamic quality sort is not driving the reported gains.","supporting_citations":[{"cited_title":", author Ma, Z","cited_arxiv_id":null,"evidence_quote":"Introduces the WMVeID863 dataset and FACENet, the fixed-TIR-primary baseline that CoEN reports outperforming by 1.6 mAP and 2.2 Rank-1."},{"cited_title":", author Liu, X","cited_arxiv_id":null,"evidence_quote":"TOP-ReID is a token-permutation multi-spectral enhancement method used as a comparison and as the TPM alternative in the CEM ablation."}],"review_version":1}