{"id":"2d433071-77d8-47fc-bac4-8cb7c5ece39c","arxiv_id":"2509.24863","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Grayscale preprocessing improves nighttime semantic segmentation robustness, but the claimed benefits of the contrast filter and of sparsification are not supported by the reported experiments.","lead":"Feeding segmentation networks grayscale or contrast-filtered versions of driving images improves nighttime robustness in experiments on Cityscapes-trained models. The paper's headline claim about 'sparse contrast representations' is not supported by the full text, which contains no sparsity experiments and shows plain grayscale often beats the contrast filter.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Contrast attribution is contradicted by the paper's own ablation: grayscale without DoG (d=0) matches or beats the contrast variant (d=5) on all three main architectures, so the robustness gain appears to come from color remapping, not contrast extraction.","rationale":"The reader identified the same weakest point: the paper assumes the gain is from DoG contrast (d=5) rather than grayscale remapping. My reading of Tables 2-3 and Appendix D confirms this. This is the single most load-bearing issue because the title, abstract, and Sec. 4.2 all claim that 'contrast-based preprocessing' and 'sparse contrast representations' are the mechanism. If the DoG term contributes nothing over color remapping, the paper's novel contribution collapses to grayscale conversion, which the related-work section already notes is known to help in low light. The sparsity claim in the abstract (70% sparsification) is also unsupported, but it is secondary: removing it would weaken the sensor-motivation, while the contrast confound directly invalidates the main explanatory claim. A paired ablation controlling d is the decisive check. Because the paper already contains the necessary comparison for three seeds and the results are consistent, I do not think the verdict should change: it should remain CONDITIONAL, requiring the authors to reframe conclusions around grayscale remapping and either substantiate or remove the sparsity claim. The provided check will settle whether the contrast component has any measurable benefit; if it does not, the condition is mandatory.","tokens_in":38371,"tokens_out":4968,"duration_ms":62623,"concrete_test":"Run a controlled comparison for the three main architectures (ResNet-50, SegFormer, InternImage): train K≥5 seeds per condition in {grayscale, color-opponency, single-color} × {d=0, d=5}, all on Cityscapes, and evaluate on Dark Zurich and ACDC Night. Compute the paired difference (d=5 − d=0) per variant with a 95% CI. If the grayscale contrast difference is ≤0 (or CI includes 0) while grayscale d=0 remains above the RGB baseline, the contrast-extraction attribution fails and the contribution reduces to grayscale conversion. Additionally, if a sparsity claim is retained, add a sparsification sweep (e.g., thresholding the preprocessed representation at 10–90%) and report the mIoU curve; the abstract's '70%' figure must be reproducible from these data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim ('biologically inspired sparse contrast representations') requires that the DoG-based contrast term in Eq. (4) is what drives the nighttime robustness gains. The ablation tables contradict this. For all three principal architectures (DeepLabv3+/ResNet-50, SegFormer, InternImage), Tables 2-3 show that grayscale with d=0 (no contrast extraction) has mIoU equal to or higher than grayscale with d=5 on Dark Zurich and ACDC Night (e.g., DeepLabv3+: 20.79 vs 18.39 on DZ; SegFormer: 21.47 vs 19.99; InternImage: 18.61 vs 17.66). Appendix D explicitly concedes that for all architectures except the smaller ResNet-18, grayscale d=0 surpasses d=5 in nighttime scenarios. Hence the improvements attributed to 'contrast-based preprocessing' are actually explained by the color-to-grayscale remapping (M in Eq. 1). The abstract's additional claim that accuracy is stable under 70% sparsification appears nowhere in the body or tables, so the 'sparse' qualifier is likewise unsupported. The load-bearing premise of the title and abstract is therefore not established by the paper's own data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a fixed, model-agnostic input preprocessing module for semantic segmentation under adverse illumination. The preprocessing combines a color reparameterization matrix M (grayscale, color-opponency, or single-color) with approximate Difference-of-Gaussians (DoG) contrast extraction, implemented by subtracting progressively box-blurred versions of the image; a depth parameter d controls the number of blur steps (Eq. (4)). Models are trained on Cityscapes and evaluated on Dark Zurich and ACDC. Across DeepLabv3+, SegFormer, and InternImage, the paper reports large mIoU gains at night for all preprocessing variants, with grayscale the strongest, while roughly preserving in-distribution accuracy. A depth search on a smaller ResNet-18 is used to select d=5, which is then applied to the main architectures. The associated abstract additionally states that accuracy remains stable under 70% sparsification of the representation, although no sparsification experiments appear in the body or appendices.","tokens_in":38721,"tokens_out":8236,"duration_ms":63332,"significance":"If the central claim were established, this would be a practically attractive result: a fixed, cheap, architecture-agnostic input transform that improves real-world nighttime robustness without training modifications. The paper has genuine strengths: it uses standard benchmarks (Cityscapes, Dark Zurich, ACDC), multiple architectures, three seeds with error bars, and a transparent ablation structure. The large and replicated gains from grayscale conversion (e.g., ACDC Night mIoU from about 8 to 20-24 for DeepLabv3+) are a useful empirical contribution even if the mechanism is less exotic than claimed. However, the significance as currently framed is weakened because the paper's own data attribute the gains to color remapping rather than to the DoG contrast term, and the 'sparse' / 70%-sparsification claim has no supporting experiments. The contribution can be salvaged by reframing the headline claim around luminance-based color remapping and either demoting the contrast mechanism to a secondary analysis or providing controlled experiments where the contrast term is evaluated on equal footing.","major_comments":[{"comment":"The claim that 'contrast-based preprocessing is particularly effective in nighttime scenarios' is contradicted by the paper's own ablations. With d=0, Eq. (4) reduces to color reparameterization without the spatial contrast term. On Dark Zurich (Table 2), DeepLabv3+ grayscale d=0 attains 20.79 mIoU versus 18.39 for grayscale d=5; SegFormer attains 21.47 versus 19.99; InternImage attains 18.61 versus 17.66. The same ordering holds on ACDC Night (Table 3: 21.87 vs 19.76, 23.93 vs 22.84, 18.70 vs 17.89). Appendix D explicitly concedes that d=0 surpasses d=5 in nighttime for all architectures except the small ResNet-18. Thus the robust improvement over RGB is explained by the color-to-grayscale remapping in Eq. (1), not by the DoG contrast term in Eq. (4). The abstract and Sec. 4.2 overstate the causal role of contrast; either the central claim must be revised or an ablation is needed that c","section":"Sec. 4.2 / Eq. (4) / Tables 2-3"},{"comment":"The depth hyperparameter d=5 is selected using the evaluation datasets themselves. Fig. 4 and the surrounding text state that d=5 was chosen by balancing performance on Cityscapes, Dark Zurich, and ACDC Night on a DeepLabv3+-ResNet-18 model, and this configuration is then used for the headline comparisons in Fig. 2 and Tables 1-3. This makes the reported 'best' contrast configuration fitted to the test benchmarks; no held-out validation or sensitivity analysis on the main architectures is provided. The authors should report results across all depths for the main architectures or select d on a validation split, so that the reader can distinguish a tuned preprocessing choice from a robust property of the method.","section":"Sec. 4.3 / Fig. 4"},{"comment":"The abstract included with the submission claims that 'the segmentation accuracy remains stable even when the contrast-based representation is sparsified by up to 70%,' and the submission title contains 'Sparse Contrast Representations.' Neither the sparsification experiment nor any explicit sparsity operation appears in the body, appendices, or Eq. (4). As described, the preprocessing produces dense contrast signals with no thresholding or pruning mechanism. If this claim is part of the paper, the experiments and the exact sparsification protocol must be added; otherwise the 'sparse' qualifier and the 70% stability statement should be removed as unsupported.","section":"Abstract vs. body"},{"comment":"The limitations discussion is internally inconsistent. Appendix D says the depth search was 'conducted on DeepLabv3+ (ResNetV1c-50), as shown in Fig. 8,' but Fig. 8 and Tables 7-12 are labelled DeepLabv3+ (ResNetV1c-18). For the ResNet-50 model in Tables 2-3, grayscale d=0 actually beats d=5 on Dark Zurich and ACDC Night, which is the opposite of the claimed illustration. This contradiction must be corrected; as written, it confuses the only architecture (ResNet-18) for which the depth d=5 contrast variant helps on nighttime data.","section":"Appendix D"}],"minor_comments":[{"comment":"The caption refers to 'InterImage'; the architecture name throughout the paper is InternImage. Please correct this typo.","section":"Fig. 2 caption"},{"comment":"The notation in Eq. (4) should be made precise: define img_i as the image after i recursive box-blur applications, state that the coefficients sum to zero, and clarify the stacking of the original and blurred images. Currently the expression mixes matrix and index notation and is hard to reproduce from the text alone.","section":"Eq. (4) and surrounding text"},{"comment":"The table formatting for the grayscale rows is confusing: rows labeled 'grayscale 0' appear under 'grayscale green bias' blocks and vice versa. Please add explicit channel counts and row labels so it is unambiguous which row corresponds to which color matrix and depth.","section":"Appendix B, Tables 1-6"},{"comment":"The full-text title is 'Vision At Night: Exploring Biologically Inspired Preprocessing For Improved Robustness Via Color And Contrast Transformations,' while the submission-level title is 'Improved Robustness from Biologically Inspired Sparse Contrast Representations.' These should be aligned, and the 'sparse' wording should be used only if supported by experiments.","section":"Title / abstract consistency"}],"recommendation":"major_revision","confidential_remarks":"The conditional reader's verdict is fair, and the skeptic's main concern lands: the paper's own ablation (Tables 2-3, Appendix D) shows that the d=0 grayscale control matches or beats the d=5 contrast variant on nighttime data for all three main architectures. The grayscale-remapping contribution is real and potentially publishable, but the current framing as a 'sparse contrast representation' result is not supported. I think a major revision is appropriate: the authors should reframe the central claim around luminance-based color remapping, either demote the DoG depth analysis or add matched experiments that can isolate the contrast term, and remove or substantiate the 70% sparsification claim. The Appendix D / Fig. 8 architecture mismatch should also be corrected before the paper is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: (1) the grayscale-preprocessing result is real and useful — it roughly triples nighttime mIoU across three architectures on Dark Zurich and ACDC — and (2) the paper's own framing is wrong. The title and abstract credit 'sparse contrast representations' and DoG-based contrast extraction, but the ablation tables show that dropping the contrast step (depth d=0) matches or beats the full pipeline (d=5) on every architecture and nighttime dataset they report. The gain comes from color-to-grayscale remapping, not from local contrast. The abstract also claims stability under 70% sparsification; there is no such experiment anywhere in the paper.\n\nWhat is genuinely new: as far as I can tell, nobody has shown that simple grayscale input consistently improves real-world nighttime semantic segmentation with this breadth (three architectures, three seeds, two external night datasets). The effect is large and the error bars are non-overlapping. The authors did the right kind of ablation, and the supplementary Appendix D concedes the d=0-vs-d=5 problem, even though the main text doesn't follow through. That's honest reporting buried in a misleading narrative.\n\nSoft spots, in proportion: the claim-attribution gap is load-bearing, not cosmetic. If the preprint survives, it should be reframed as a grayscale-preprocessing result, with the contrast module presented as a variant that doesn't help at night. The d=5 depth was selected on the evaluation datasets (using ResNet-18) and then fixed for the main runs, which is test-set tuning of a free parameter; that complicates the headline numbers. There are no significance tests, though the effect sizes are large. The sparsity sentence in the abstract is simply unsupported and should be removed or backed by new experiments. The citation list is heavy on the authors' own group, but that is not disqualifying on its own.\n\nWho should read it: anyone working on semantic segmentation robustness in autonomous driving, and anyone who wants a clean example of why you must ablate the individual components of a 'biologically inspired' pipeline. The empirical core deserves a serious referee and a major revision — not a desk reject. If they fix the framing and remove the unsupported claim, this becomes a small but solid contribution.\n\nMy recommendation: send it to peer review, expecting a revised version that rewrites the title, abstract, and conclusions to match the data, plus code release. I would not want to see it accepted in its current state.","headline":"Useful grayscale-preprocessing result undermined by a misattributed contrast mechanism and an unsupported sparsity claim; deserves serious revision, not rejection.","tokens_in":39169,"tokens_out":4950,"would_cite":true,"duration_ms":34564,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A fixed, retina-inspired preprocessing module improves semantic segmentation in night and adverse weather without retraining.","keywords":["semantic segmentation","robustness","nighttime","Difference of Gaussians","color-opponency","grayscale preprocessing","distribution shift","biological vision"],"falsifier":"Reproduce the paper's Cityscapes training with grayscale d=0, grayscale d=5, and RGB on Deeplabv3+, SegFormer, and InternImage, and test on Dark Zurich and ACDC Night. If grayscale d=0 matches or beats grayscale d=5 for all three architectures on both datasets, the specific claim that contrast extraction improves nighttime robustness is falsified; also check whether any section reports the 70% sparsification result claimed in the abstract.","tokens_in":38318,"feed_emoji":"🌙","tokens_out":7065,"duration_ms":47933,"temperature":0.7,"pith_summary":"This paper argues that replacing raw RGB input with a fixed, biologically motivated preprocessing transform can make semantic segmentation models robust to lighting and weather shifts, without changing the network or training procedure. The transform combines color reparameterization (grayscale, color-opponency, or single-color) with Difference-of-Gaussians (DoG) local contrast extraction at a fixed depth. Trained on Cityscapes, the preprocessed models keep in-distribution mIoU roughly on par with RGB baselines while substantially improving mIoU on Dark Zurich and the night split of ACDC. A careful reader would care because the approach is model-agnostic and lightweight, so it could be pushed into imaging sensors as a near-acquisition operation. The paper also claims that the contrast representation tolerates up to 70% sparsification, pointing toward transmission-friendly pipelines.","feed_headline":"Retina-style preprocessing lifts nighttime segmentation by 14 points","feed_subtitle":"A fixed, model-agnostic input transform keeps day accuracy and boosts night segmentation by up to 14 points.","key_machinery":"The central object is the fixed preprocessing module defined by img_contrast = M·img_original − Σ_{i=1..d} (1/d)·M·img_i, where M is a 3×3 color-reparameterization matrix (grayscale, color-opponency, or identity) and img_i are progressively box-blurred versions of the image. With d=5, the module approximates a Difference-of-Gaussians (DoG) center-surround kernel, acting like retinal center-surround cells by keeping local contrast while discarding absolute luminance. The depth d controls how much spatial context the contrast subtraction uses; d=0 removes contrast and leaves only the color transform, serving as the control. This module is what carries the argument: because it is frozen and app","core_discovery":"On its own terms, the paper's central discovery is that the way an image is represented before it enters a segmentation network is itself a robustness lever. The preprocessing module first maps RGB through a fixed linear color transform M, then subtracts progressively blurred versions of the transformed image (box blur applied d=5 times) to produce a local-contrast signal that discards absolute luminance. Across three architectures, this fixed transform preserves Cityscapes performance (mIoU within a few points of RGB baseline) while raising mIoU on Dark Zurich and ACDC Night from single digits to roughly 18-23, a large relative gain. Grayscale variants (with or without contrast) perform bes","pith_inferences":["The paper's own ablation data suggest the 'contrast' component may be secondary: in Tables 2 and 3, grayscale with d=0 (no contrast) matches or beats grayscale with d=5 on Dark Zurich and ACDC Night for Deeplabv3+, SegFormer and InternImage. If that pattern generalizes, the practical takeaway would be that color removal, not center-surround contrast, drives night robustness.","A natural testable extension is to apply the module to other dense tasks such as depth estimation and detection, where the luminance-versus-contrast attribution can be controlled more cleanly than in real datasets.","The 70% sparsity claim appears in the abstract but not in the visible experimental sections; verifying it with explicit sparsification curves would determine whether the transmission-aware imaging-sensor story is supported.","Because optimal depth varies with architecture and color transform, a learned or self-adaptive depth parameter seems a promising follow-up, potentially improving on the fixed d=5 choice."],"forward_implications":["If the claim holds, robustness to night and weather can be improved by changing only the input representation, leaving architectures and training pipelines untouched.","The preprocessing is cheap enough (two or three fixed convolutional layers) to be embedded in imaging sensors, so cameras could deliver task-ready robust inputs directly.","Grayscale-based inputs, in particular, can serve as a strong and simple baseline for nighttime semantic segmentation, since all grayscale variants beat RGB at night across tested architectures.","The reported tolerance to 70% sparsification suggests the contrast representation could be compressed or thresholded near acquisition, reducing transmission bandwidth in sensor networks.","Color-opponent contrast helps in weather conditions where low light is not the main issue, indicating that the optimal representation depends on the dominant corruption."],"fun_headline_variants":["Retina-inspired preprocessing lifts night segmentation up to 14 points","Sparse contrast transform preserves day accuracy, boosts night mIoU","Fixed preprocessing mimics retina: robust to illumination shifts","Sparsify 70%: contrast coding keeps segmentation stable at night"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim stands on the assumption that the robustness gains come from the biologically inspired contrast-extraction module rather than from the simpler color-to-grayscale remapping; the paper's ablations show grayscale without contrast (d=0) often equals or beats the contrast version, and the abstract's 70% sparsity-stability figure is not backed by experiments in the visible text.","fun_headline_variants_meta":{"raw":{"variants":["Retina-inspired preprocessing lifts night segmentation up to 14 points","Sparse contrast transform preserves day accuracy, boosts night mIoU","Fixed preprocessing mimics retina: robust to illumination shifts","Sparsify 70%: contrast coding keeps segmentation stable at night"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":2981,"prompt_tokens":718,"completion_tokens":2263,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":2203}},"tokens_in":462,"tokens_out":2263,"duration_ms":14608,"temperature":1.0,"reasoning_tokens":2203,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:37:34.400515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the paper's Cityscapes training with grayscale d=0, grayscale d=5, and RGB on Deeplabv3+, SegFormer, and InternImage, and test on Dark Zurich and ACDC Night. If grayscale d=0 matches or beats grayscale d=5 for all three architectures on both datasets, the specific claim that contrast extraction improves nighttime robustness is falsified; also check whether any section reports the 70% sparsification result claimed in the abstract.","supporting_citations":[],"review_version":1}