{"id":"d89bfc0c-5e4b-4171-9dde-62e94c0c8450","arxiv_id":"1908.05142","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stream person re-identification network that trains RGB and grayscale branches and fuses them with a joint branch improves mAP on four benchmarks, with grayscale input contributing most on the hardest dataset.","lead":"This paper finds that adding a grayscale copy of each person image helps a two-stream neural network identify people, because color alone can mislead the model. The proposed GreyReID framework merges color, grayscale, and fused features to improve person re-identification benchmarks, but the gains over the same architecture without grayscale are small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 8's decisive RGB-grey vs RGB-RGB gains (0.3–1.4 mAP) come from single unseeded runs; the paper's own Table 4 vs Table 8 numbers differ by a similar magnitude, so the central claim needs error bars or seeds.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: Table 8's RGB-grey vs RGB-RGB differences are small, single-run, and unreplicated. My read of the full text reinforces this and adds a concrete internal inconsistency: the configuration that should match the final method (Table 4, Two-part, Grey+RGB+Joint) reports different numbers from Table 8's RGB-grey pair on every dataset. This suggests either run-to-run variance of the same magnitude as the claimed effect, or an unreported difference in setup, and it makes the absence of error bars and seeds more consequential. I do not see a different, more fundamental flaw: the two-stream architecture itself is supported by consistent, larger gains over the baseline, the ablations are logically structured, and the comparison isolates the greyscale input by replacing it with RGB while keeping the rest of the framework fixed. The correct response is therefore to hold the paper to its conditional status: the greyscale-specific improvement must be demonstrated with multiple seeds and significance testing before the central claim is accepted. This does not change the reader's CONDITIONAL verdict, so I mark it UNCHANGED.","tokens_in":20463,"tokens_out":6895,"duration_ms":64725,"concrete_test":"Re-run the Table 8 RGB-RGB and RGB-grey configurations and the Table 4 Two-part Grey+RGB+Joint configuration on MSMT17 (and ideally Market1501) with 5 random seeds each, keeping all hyperparameters fixed; report per-seed mAP/Rank-1 with mean and std. The greyscale benefit is supported only if the paired RGB-grey minus RGB-RGB difference is consistently positive across seeds and exceeds the seed-level std; also verify whether any configuration matches Table 4's numbers, and if not, identify the omitted difference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that greyscale input improves Re-ID rests entirely on Table 8, where RGB-grey beats RGB-RGB by 0.5/0.8/0.3/1.4 mAP on Market1501/Duke/CUHK03/MSMT17, with Rank-1 gains of 0.5/0.5/0.9/1.3. Each configuration is run once, with no standard deviations, significance tests, or seeds reported. Given that typical Re-ID training noise is on the order of 1 mAP, these margins — especially the 0.3 mAP CUHK03 gain — are within run-to-run variance. The concern is sharpened by an internal inconsistency: Table 4's Two-part 'Grey+RGB+Joint' row should be the same configuration as Table 8's 'RGB-grey pair' (two-part framework, concatenated features), but the numbers differ on every dataset (e.g., Market1501 85.3 vs 85.6 mAP; Duke 75.9 vs 76.5; MSMT17 54.5 vs 55.0). This either means the two tables come from different runs or from unreported differences in feature dimensions or hyperparameters; both possibilities undermine the single-run comparison. If the RGB-grey vs RGB-RGB deltas are within the same noise that separates Table 4 from Table 8, the paper's novel conclusion that greyscale information improves Re-ID is not established, even though the two-stream framework's larger gains over Baseline (78.5→85.6 mAP on Market1501) may be real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GreyReID, a two-stream deep framework for person re-identification that jointly exploits RGB and grayscale images. RGB images are converted to grayscale using a standard luminance formula; the two modalities are processed by separate ResNet-50 branches, fused into a joint branch with part-based pooling, and the features of all three branches are concatenated for retrieval. Training uses per-branch and global losses combining cross-entropy and hard-mining triplet loss. Experiments on Market1501, DukeMTMC-reID, CUHK03-NP, and MSMT17 report gains over a single-stream baseline, comparisons with state-of-the-art methods, and an ablation showing that the RGB-grey input pair outperforms an RGB-RGB pair in the same architecture. The central claim is that RGB and grayscale information are complementary for person Re-ID.","tokens_in":20790,"tokens_out":5740,"duration_ms":49186,"significance":"If the empirical claim holds, the paper makes a simple and practical contribution: it identifies color over-fitting as a failure mode in Re-ID and shows that a conversion-based grayscale stream, requiring no extra annotations or learned generation, can improve accuracy when fused with the RGB stream. The experimental design is commendable: the RGB-RGB pair condition isolates the effect of the grayscale input by keeping the architecture and training pipeline fixed, and the results are directionally consistent across four benchmarks. The paper is also transparent in reporting ablations over loss components, fusion schemes, feature dimensions, backbones, and conversion methods, and it does not rely on fitted constants or circular derivations. However, the magnitude of the central effect (0.3–1.4 mAP) is small relative to typical training noise, and the paper provides no error bars, seeds, or significance tests, which currently limits the strength of the conclusion.","major_comments":[{"comment":"The central claim that RGB-grey input is complementary to RGB input rests entirely on the comparison between the 'RGB-RGB pair' and 'RGB-grey pair' rows in Table 8, where the gains are 0.5/0.8/0.3/1.4 mAP and 0.5/0.5/0.9/1.3 Rank-1 on Market1501/Duke/CUHK03/MSMT17. The paper reports only single-run results with no standard deviations, no seeds, and no significance tests, and these margins are within the run-to-run variance commonly observed in deep Re-ID training. As written, the evidence does not rule out the possibility that the apparent improvement from grayscale is noise; the authors should report mean and standard deviation over at least five independent runs with fixed seeds and apply a paired significance test to the RGB-grey versus RGB-RGB comparison.","section":"Section 4.5, Table 8"},{"comment":"The manuscript contains an internal inconsistency that bears directly on the central claim. The 'Two-part Grey+RGB+Joint' row in Table 4 should describe the same configuration as the 'RGB-grey pair' row in Table 8 (two-part architecture with all three branch features concatenated), yet the reported numbers differ on every dataset: Market1501 85.3 versus 85.6 mAP, Duke 75.9 versus 76.5, CUHK03 69.6 versus 69.9, MSMT17 54.5 versus 55.0. In addition, Table 12 lists '256+512+512=1280' with 86.0 mAP on Market1501, while the main results report 85.6 for the same nominal feature configuration. The authors need to state explicitly whether these are different runs, different hyperparameters (e.g., λ or feature dimensions), or different data splits, and to make the configuration used for each reported number unambiguous; the observed gaps are of the same order as the claimed RGB-grey advantage and currently undermine the reliability of the comparison.","section":"Section 4.4, Table 4; Section 4.5, Table 8"}],"minor_comments":[{"comment":"The table header contains the typo 'Mehods'; please change to 'Methods'. Also, 'CHUK03' in the captions of Tables 4, 6, and 8 should be 'CUHK03'.","section":"Section 4.3, Table 1"},{"comment":"The phrase 'This can be excerpted because' should read 'This can be expected because'.","section":"Section 4.2"},{"comment":"In the discussion of Table 9, 'smilar' and 'RBG' should be 'similar' and 'RGB'; additionally, the sentence 'as a RBG image does' is grammatically unclear.","section":"Section 4.5"},{"comment":"The heading 'The different between the proposed method and the cross-modal method' should be 'The difference between the proposed method and cross-modal methods'.","section":"Section 3.3"},{"comment":"It would help reproducibility to state whether random erasing is applied before or after the RGB-to-grayscale conversion, and whether the grayscale branch receives three replicated channels (as mentioned in Section 3.1) with the same normalization statistics as the RGB branch.","section":"Section 4.1"},{"comment":"The caption of Fig. 3 is confusing: the parenthetical color descriptions for the bars (orange, grey, blue) do not match the visible bar labels in the figure, and the sentence 'Blue bars represent using RGB and greyscale images to train and test the ResNet-50 model, respectively' should be reworded to identify which model is evaluated on which modality.","section":"Section 4.5, Fig. 3"},{"comment":"The state-of-the-art comparison tables list mostly methods from 2018 and earlier; to support the 'state-of-the-art' claim, the authors should add contemporary baselines or soften the wording in Section 4.3.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clean idea and a well-structured ablation, but the central claim is currently under-supported statistically. I would ask the editor to require multi-seed experiments and reconciliation of Table 4, Table 8, and Table 12 before acceptance; if the authors cannot provide error bars, the claim that grayscale helps should be retracted or substantially weakened. The paper's fit with TOMM is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the stress-test note lands. The paper's central claim—greyscale input helps Re-ID—rests on Table 8, and those gains (0.3–1.4 mAP) are within ordinary run-to-run noise for this task. The paper reports single runs, no seeds, no error bars. Worse, the configuration called \"Two-part Grey+RGB+Joint\" in Table 4 and the \"RGB-grey pair\" in Table 8 look like the same setup but differ on every dataset (85.3 vs 85.6 mAP on Market1501, 75.9 vs 76.5 on Duke, 69.6 vs 69.9 on CUHK03, 54.5 vs 55.0 on MSMT17). That discrepancy tells me either the tables come from different runs or there is an unreported change in hyperparameters. Either way, the precision needed to support the headline is not there.\n\nThat said, this is not a sloppy paper. The ablations are thoughtful: the RGB-RGB pair is the right control; they test HSV, greyscale-as-augmentation, different fusion schemes, loss components, and feature dimensions. The direction is consistent across four datasets. The overall framework gets a large, credible gain over baseline (Market mAP 78.5→85.6), and that gain is not in question. What is in question is only how much of it comes from the greyscale stream specifically versus the two-stream architecture plus joint training.\n\nNovelty is real but modest. The combination of two ResNet-50 streams, element-wise fusion, part pooling, and cross-entropy-plus-triplet loss is standard; the greyscale input is the new variable. No code or data is released, though the method is described well enough to reimplement. The citation pattern looks fine, including self-citations to prior Re-ID work from the same group.\n\nMy overall read: this paper is worth taking seriously and worth sending to reviewers. The reviewers should ask for multi-seed runs with standard deviations, a reconciliation of Tables 4 and 8, and code. The conceptual claim that greyscale and RGB are complementary is plausible and the evidence is suggestive, not conclusive. The paper's own Table 13, where greyscale as plain augmentation hurts, makes the two-stream result interesting even if the margins are small. This is useful reading for anyone working on input representations or color robustness in person Re-ID.","headline":"The RGB-grey versus RGB-RGB comparison is the one genuinely new piece here, and the paper's ablations are careful, but the central claim rests on single-run margins that are within Re-ID training noise.","tokens_in":21331,"tokens_out":2216,"would_cite":true,"duration_ms":22509,"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":"A grayscale twin stream beats an RGB-only twin for person re-identification.","keywords":["person re-identification","two-stream network","grayscale images","color over-fitting","feature fusion","triplet loss","Market1501","MSMT17"],"falsifier":"Rerun the Table 8 comparison ten times with different random seeds on MSMT17, where the reported RGB-grey advantage is largest (1.4 mAP), and check whether the paired difference is consistently positive across runs; if the advantage is smaller than the run-to-run standard deviation, the claim that grayscale input improves person re-identification is not supported.","tokens_in":20257,"feed_emoji":"👤","tokens_out":6496,"duration_ms":59837,"temperature":0.7,"pith_summary":"The paper argues that color is a double-edged cue in person re-identification: most false matches in a ranking list share color with the query, so RGB models over-fit to color and miss structure and texture. It claims that RGB and grayscale views are complementary, because a model trained on grayscale images recovers many hard cases that an RGB model misses. To exploit this, GreyReID runs the same image through a two-stream ResNet-50, one stream on RGB and one on grayscale, merges the two at a joint branch, and concatenates all three features as the final descriptor. On Market1501, DukeMTMC-reID, CUHK03-NP, and MSMT17, the RGB+grayscale input pair consistently beats the same architecture using two RGB streams, and the full method outperforms the compared state of the art. If true, the finding offers a parameter-free way to reduce color over-fitting with no extra data or models.","feed_headline":"Grayscale stream beats RGB-RGB for person re-ID","feed_subtitle":"A two-stream network that adds a grayscale copy lifts retrieval accuracy on all four benchmarks, most on MSMT17.","key_machinery":"The load-bearing mechanism is a two-stream deep network whose RGB and grayscale branches are fused before pooling. Each branch is a pretrained ResNet-50; the grayscale input is generated by the weighted conversion $Grey(i,j)=0.299R(i,j)+0.587G(i,j)+0.114B(i,j)$ and replicated to three channels. The two branches' last convolutional tensors are combined by element-wise addition into a joint tensor, which is split into two non-overlapping parts before global average pooling, and the three resulting 256-, 512-, and 512-dimensional features are concatenated into a 1280-dimensional descriptor. Each branch is supervised independently by a cross-entropy loss plus a hard-mining triplet loss, and the concatenated feature receives a global loss, so the framework jointly learns to trust color, structure, and their interaction.","core_discovery":"The central claim is that removing color information helps person re-identification when the grayscale view is combined with the RGB view through an explicitly fused architecture, rather than used as ordinary data augmentation. The paper supports this with three observations: RGB models lose over 44% mAP when tested on grayscale images, a grayscale-trained model recovers much of that gap, and the two models' feature response maps attend to different regions. Its network feeds RGB and grayscale versions of each image into two ResNet-50 streams, sums the last-convolution tensors into a joint branch, and concatenates the branch features; each branch is trained with its own cross-entropy plus triplet loss, and a global loss tunes the concatenation. The decisive comparison in Section 4.5 uses identical settings and reports that RGB-grey inputs outperform RGB-RGB inputs on all four benchmarks, by 0.5, 0.8, 0.3, and 1.4 mAP respectively. Greyscale images used as plain data augmentation, by contrast, degrade or fail to improve performance, which the paper takes as evidence that the fusion architecture is what extracts the value.","pith_inferences":["Our inference: the consistent RGB-grey advantage over RGB-RGB suggests the grayscale stream acts as a regularizer against color over-fitting, and the same idea could be tested in other retrieval tasks where color is dominant but noisy, such as vehicle re-identification.","Our inference: since the paper runs each configuration once, practitioners should rerun the RGB-grey versus RGB-RGB comparison with multiple seeds before treating the 0.3–1.4 mAP gaps as deployment guarantees; the reported pattern is directionally consistent but variance is unreported.","A testable extension the paper does not explore is replacing the grayscale stream with other color-invariant views, such as Lab lightness or an adversarially color-perturbed stream, to see whether the benefit comes specifically from grayscale or from any view that suppresses color.","The two-stream idea could be adapted to RGB-infrared cross-modality Re-ID by using the grayscale branch as a shared intermediate representation, although the paper explicitly distinguishes its same-modality task from cross-modality matching."],"forward_implications":["A cheap preprocessing step—converting RGB images to grayscale in each batch—adds a complementary stream that improves mAP and Rank-1 on all four benchmarks, with the largest gain on the largest dataset, MSMT17.","The per-branch and global losses are necessary; removing branch losses drops mAP substantially (85.6 to 79.1 on Market1501), showing the architecture benefits from independent supervision of each modality.","Simply mixing grayscale images into the training set as augmentation does not help (Table 13) and often hurts; the fusion into a joint branch is what captures the complementarity.","The framework generalizes across backbones: switching ResNet-50 to ResNet-101 raises mAP on all four datasets, for example from 55.0 to 59.1 on MSMT17.","Because the grayscale conversion is a fixed pixel operation that requires no extra training data or generative models, the approach can be attached to existing part-based or attention-based Re-ID backbones."],"supporting_citations":[{"why":"Provides the PCB strong baseline and the ranking-list examples that motivate the color over-fitting observation.","marker":"[29]"},{"why":"ResNet-50 is the backbone of both branches and of the baseline model.","marker":"[10]"},{"why":"Supplies the hard-mining triplet loss used in every branch and the global loss.","marker":"[11]"},{"why":"Defines the Market1501 benchmark and the mAP/Rank-1 evaluation protocol.","marker":"[45]"},{"why":"Defines the DukeMTMC-reID benchmark used for comparison.","marker":"[24]"},{"why":"Establishes the CUHK03-NP train/test split used for CUHK03 evaluation.","marker":"[51]"},{"why":"Defines the MSMT17 benchmark, the dataset where the RGB-grey advantage is largest.","marker":"[35]"},{"why":"ImageNet pretraining initializes the ResNet backbones and is load-bearing for the learned feature quality.","marker":"[5]"}],"fun_headline_variants":["Fusing grayscale with RGB lifts person re-ID","Why removing color helps person re-ID","Grayscale+RGB fusion tops single-color for re-ID","Add grayscale, not color, to boost re-ID","Colorblind stream improves person re-ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small, consistent improvements of the RGB-grey pair over the RGB-RGB pair are genuine training effects and not run-to-run noise, because each configuration is evaluated a single time without reported variance.","fun_headline_variants_meta":{"raw":{"variants":["Fusing grayscale with RGB lifts person re-ID","Why removing color helps person re-ID","Grayscale+RGB fusion tops single-color for re-ID","Add grayscale, not color, to boost re-ID","Colorblind stream improves person re-ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1427,"prompt_tokens":1091,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":707,"tokens_out":336,"duration_ms":3318,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:21:45.565906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Table 8 comparison ten times with different random seeds on MSMT17, where the reported RGB-grey advantage is largest (1.4 mAP), and check whether the paired difference is consistently positive across runs; if the advantage is smaller than the run-to-run standard deviation, the claim that grayscale input improves person re-identification is not supported.","supporting_citations":[{"cited_title":"In The European Conference on Computer Vision (ECCV) (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the PCB strong baseline and the ranking-list examples that motivate the color over-fitting observation."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet-50 is the backbone of both branches and of the baseline model."},{"cited_title":"In defense of the triplet loss for person re-identification","cited_arxiv_id":null,"evidence_quote":"Supplies the hard-mining triplet loss used in every branch and the global loss."},{"cited_title":"Scalable person re-identification: A benchmark","cited_arxiv_id":null,"evidence_quote":"Defines the Market1501 benchmark and the mAP/Rank-1 evaluation protocol."},{"cited_title":"Performance measures and a data set for multi-target, multi-camera tracking","cited_arxiv_id":null,"evidence_quote":"Defines the DukeMTMC-reID benchmark used for comparison."},{"cited_title":"Re-ranking person re-identification with k-reciprocal encoding","cited_arxiv_id":null,"evidence_quote":"Establishes the CUHK03-NP train/test split used for CUHK03 evaluation."},{"cited_title":"Person transfer gan to bridge domain gap for person re-identification","cited_arxiv_id":null,"evidence_quote":"Defines the MSMT17 benchmark, the dataset where the RGB-grey advantage is largest."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"ImageNet pretraining initializes the ResNet backbones and is load-bearing for the learned feature quality."}],"review_version":1}