{"id":"35c6d505-eb90-4287-8c71-08688d580346","arxiv_id":"2509.06116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"CARDIE clusters images by luminance and dominant hue, and using those clusters to oversample minority classes improves PSNR for tone mapping and denoising compared to semantic clustering baselines.","lead":"The paper introduces CARDIE, a simple clustering method that groups photos by brightness and color instead of by what objects they contain, and shows that training image-enhancement models on rebalanced versions of these groups improves tone mapping and denoising. A specialist might read it because it offers a cheap, label-free way to find and fix underrepresented cases in image enhancement datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4's two-parameter Naka-Rushton assumption is the load-bearing link: if it does not capture a local operator like HDRNet, the KS-based relevance comparisons in Sec. 3 can favor CARDIE for artifacts of misfit rather than for enhancement relevance.","rationale":"I read the paper as claiming that CARDIE's low-level descriptors define clusters aligned with how HDRNet and NAFNet transform images, and that oversampling these clusters improves tone-mapping and denoising performance. The Sec. 3 evidence for alignment is the KS indicator built on per-image fits of Eq. 4. That is the most load-bearing assumption because it enters every pairwise relevance judgment. Eq. 4 is a two-parameter global curve, whereas HDRNet is a local operator; the paper reports only a few fit examples and no aggregate goodness-of-fit. If Eq. 4 is misspecified, fitted gamma/mu distributions can separate CARDIE clusters by construction, since CARDIE separates on the luminosity dimension to which Eq. 4 is sensitive, and the 'more relevant than semantic' claim becomes an artifact. The proposed empirical-curve test would settle this: recompute the indicator matrices with a flexible per-image transfer estimate. If separation disappears, the relevance claim is unsupported; if it survives, the concern is answered. The denoising experiment's extra-noise injection into CARDIE minority classes and the absence of error bars are additional concerns, but the functional-form assumption is primary. Since the reader's conditional verdict already identifies this assumption and asks for validation, I do not recommend changing the verdict.","tokens_in":9672,"tokens_out":9288,"duration_ms":87030,"concrete_test":"Recompute the cluster-relevance analysis without assuming Eq. 4: for each image pair, estimate the empirical per-pixel input-output luminance relation using a flexible estimator (e.g., a 256-bin lookup table or a spline with 5 degrees of freedom), then repeat the KS-based indicator matrices of Eqs. 6-8 for CARDIE, MIT5K-descriptor, and ResNet-Places365 clusters. If, under the flexible estimator, CARDIE no longer separates from the semantic clusters, the claimed relevance is an artifact of the assumed functional form. Also report the distribution of per-image R^2 for Eq. 4 fits; if a substantial fraction of HDRNet fits have low R^2, the metric is not trustworthy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Sec. 3 claim that CARDIE clusters are more relevant to image enhancement than semantic clusters is built entirely on per-image fits of Eq. 4 and on KS tests over the fitted gamma/mu distributions (Eqs. 6-8). For HDRNet, a local tone mapper, output luminance at a pixel depends on local context, so no global curve L_pp = L_i^gamma / (L_i^gamma + mu^gamma) is the true generator. The paper shows only a few visually satisfactory fits (Fig. 2) and reports no per-image goodness-of-fit or residual analysis over the roughly 5000 images. If the fits are systematically poor, or if gamma/mu are poorly identified, the KS differences between CARDIE clusters can reflect fitting artifacts and parameter instability rather than genuinely different enhancement transformations. Because CARDIE clusters are built from the same luminosity variable that drives Eq. 4's parameters, the test is also biased toward finding separation. A secondary but compounding issue is the denoising experiment: extra noise was deliberately injected into CARDIE-defined minority classes before showing that CARDIE-based oversampling helps, which can inflate the +1.18 dB gain. The functional-form assumption is the primary concern because it underpins the pairwise relevance metric for every operator and dataset.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces CARDIE, an unsupervised clustering algorithm that maps each image to binary descriptors for luminosity level and dominant hue intervals, then clusters the resulting descriptor vectors with HDBSCAN. To evaluate cluster usefulness, the authors propose fitting the Naka-Rushton curve of Eq. (4) to each input/output image pair and testing, via pairwise KS tests, whether the fitted gamma and mu distributions differ across clusters (Eqs. 6-7); for denoising they use the change in average local variance (Eqs. 5 and 8). They compare CARDIE with semantic clustering baselines derived from MIT5K manual labels and ResNet50-Places365 attributes on the MIT5K (D5K) and HDR+ datasets, and report that oversampling CARDIE minority clusters improves PSNR for HDRNet tone mapping (+0.22 dB full, +0.60 dB minority) and NAFNet denoising (+1.18 dB full, +2.73 dB minority) relative to baseline.","tokens_in":9973,"tokens_out":6472,"duration_ms":56248,"significance":"If the claims hold, the paper offers a simple, training-free alternative to semantic clustering for image-enhancement dataset analysis and rebalancing, with a public implementation as a reproducibility asset. The descriptor construction is transparent and computationally light, and the comparison with two semantic baselines is a useful step for the community. The reported PSNR gains are practically relevant for tone mapping and denoising pipelines. However, the evidence rests on an unvalidated global functional form, on multiple-testing issues, and on a denoising experiment with an injected-noise confound, so the significance is conditional on the requested revisions.","major_comments":[{"comment":"The cluster-relevance comparison rests entirely on the assumption that each tone mapping operator can be summarized by the two-parameter Naka-Rushton curve L^pp = L^i^gamma / (L^i^gamma + mu^gamma). For a local operator such as HDRNet, output luminance at a pixel depends on local context, so this global functional form is not the data-generating process. The paper reports only visual fit examples in Fig. 2 and provides no per-image residuals, R^2 values, or parameter-identifiability statistics over the roughly 5,000 images. If the fits are systematically poor, or if gamma and mu are unstable, the KS differences between CARDIE clusters in Fig. 5 can reflect fitting artifacts rather than enhancement-relevant differences. The authors should provide quantitative goodness-of-fit diagnostics and a robustness check with an alternative functional form or with a direct comparison of input/output luminance histograms.","section":"§2.2, Eq. (4), and §3.2"},{"comment":"CARDIE's descriptors are exactly the luminosity level and hue that determine the fitted gamma and mu in Eq. (4), so the cluster difference indicator I^T_l,a is biased toward finding separation for CARDIE. For example, low-luminosity clusters will by construction produce different mu from high-luminosity clusters for any tone mapper that shifts mean luminance. This construction bias does not by itself invalidate the resampling results, but it weakens the comparative claim that CARDIE is more relevant than semantic clustering, because the relevance metric shares the same variables that CARDIE clusters on. The authors should either use a metric not defined through the clustered variables, such as direct perceptual or pixel-wise error of the enhancement output, or explicitly justify why this construction bias does not drive the Sec. 3 conclusions.","section":"§2.1-§2.2, cluster difference indicator Eq. (7)"},{"comment":"Pairwise KS tests are performed for every cluster pair and a p-value threshold of 0.05 is used without any multiple-testing correction. With 12 clusters on D5K there are 66 pairwise comparisons, so several false 'statistically different' results are expected under the null hypothesis. The cluster difference indicators in Figs. 5-6 and the conclusions drawn from them should be recomputed with adjusted p-values, for example using Benjamini-Hochberg correction, or the authors should justify the uncorrected alpha level and report the number of tests performed.","section":"§2.2, Eqs. (6)-(8), Figs. 5-6"},{"comment":"The denoising experiment is confounded by the deliberate injection of approximately twice the noise into the two least represented classes of D5K before the oversampling experiment. Because the same classes are then oversampled by CARDIE, the +1.18 dB full-dataset and +2.73 dB minority-class gains in Tables 5 and 6 can be explained by the extra noise added to those classes rather than by CARDIE's cluster relevance. A control experiment with uniform noise, or with the same noise-injection protocol applied to random and ResNet oversampling, is required to attribute the gain to the clustering method.","section":"§3.1 and §4, Tables 5-6"},{"comment":"The PSNR differences are reported from what appears to be a single training run per condition, with no standard deviations, error bars, or repeated-seed results. The claimed improvements (0.22 dB and 1.18 dB) are small relative to typical run-to-run variation in trained HDRNet and NAFNet models. At minimum, the authors should report mean and standard deviation over three or more seeds, and ideally a paired significance test across runs.","section":"§4, Tables 5-6"}],"minor_comments":[{"comment":"There are typos that should be corrected: 'Nice Reseach Center' should be 'Nice Research Center', and 'ground thruth' in the Fig. 10 caption should be 'ground truth'.","section":"Author affiliation and Fig. 10 caption"},{"comment":"Equation (3) mixes image-level and pixel-level notation: L(x^pp_j, p) = A(L(x^i_j), p) is hard to parse. Please clarify that A is applied to the luminance channel and what p indexes in the operator argument.","section":"§2.2, Eq. (3)"},{"comment":"The text does not specify whether the oversampling for the denoising experiment uses cluster labels computed on D5K or on D_noisy; Tables 2 and 3 show different cluster structures, so this choice matters for reproducibility. Please state explicitly which cluster labels are used for each task.","section":"§4, oversampling protocol"},{"comment":"The noise model x_noisy ~ N(x_gt, a_j x_gt + b_j) does not state whether pixel values are in [0,1] or [0,255]; the reported ranges for a_j and b_j are otherwise ambiguous.","section":"Eq. (9)"},{"comment":"The claim that the results are 'not too sensitive' to P=100 is not accompanied by any sensitivity analysis. Please add a short experiment varying P or state the range of values tested.","section":"§2.2, sensitivity to P"},{"comment":"The figure captions should explicitly define the color code (green = I_l,a = 1, red = I_l,a = 0) rather than leaving it to the main text.","section":"Figs. 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for a computer vision journal and the public code release is a genuine asset. The main risk is that the central comparison relies on an unvalidated parametric model and on an experiment with an injected-noise confound. I would invite a revision with the requested controls rather than reject, but I would not accept the paper in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. This paper proposes CARDIE, an unsupervised clustering method that labels images by low-level luminance and hue descriptors and clusters them with HDBSCAN. The new part is combining these simple descriptors with a Naka-Rushton-based relevance metric to guide dataset resampling for image enhancement. It's a genuinely useful idea, and the authors compare it against semantic clustering (ResNet Places365 and MIT5K labels) and release their code.\n\nWhat the paper does well: the pipeline is refreshingly low-tech, the comparisons are sensible, and the downstream resampling experiments on tone mapping (HDRNet) and denoising (NAFNet) are the right way to test practical value. The HDR+ extension in the appendix is a good robustness check. If the PSNR gains are real, oversampling CARDIE minority clusters is a cheap win for practitioners.\n\nNow the soft spots, in rough order of concern. First, the headline PSNR numbers (Tables 5-6) come from a single training run per condition, with no error bars or seed variation. A 0.22 dB full-set gain and a 1.18 dB denoising gain could easily move within run-to-run noise for these models. Second, the cluster relevance indicator (Eq. 7-8) runs many pairwise KS tests at alpha=0.05 without multiple-testing correction, and the OR-logic (gamma or mu) makes differences easier to flag. This likely overstates how distinctive the clusters are. Third, the synthetic denoising dataset is built by injecting extra noise into CARDIE-defined minority classes; that makes the oversampling benefit partly self-fulfilling, though it's a fair test of rebalancing if the test set has the same distribution. Fourth, and this is the load-bearing assumption, Eq. 4 fits a two-parameter global Naka-Rushton curve to each image under a tone mapper. For a local operator like HDRNet, that functional form is misspecified, and the paper shows only a few fits with no residual analysis across the roughly 5000 images. If gamma and mu are poorly identified, the KS differences between CARDIE clusters could reflect fitting artifacts rather than genuine differences in enhancement. There's also an inherent circularity: CARDIE clusters by luminosity, and a global tone mapper's fitted parameters depend on luminosity, so separation is partly baked in.\n\nThat said, none of this is fatal. The resampling result is a concrete downstream outcome independent of the KS metric, and the method is simple enough to verify quickly. I'd send it to peer review with a request for: multiple seeds with mean and std, proper multiple-testing handling, residual diagnostics for Eq. 4, and a less contrived noise experiment (or a clear justification for the construction). The paper targets researchers in image enhancement and dataset analysis; after revision it could be a solid practical contribution.","headline":"A useful, low-tech clustering method for rebalancing image enhancement datasets, but the relevance metric rests on a shaky functional fit and the PSNR gains need error bars.","tokens_in":10507,"tokens_out":4409,"would_cite":false,"duration_ms":39128,"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":"The paper claims that unsupervised clustering by luminosity and hue, rather than by semantic content, yields the clusters that matter for image enhancement, and that oversampling those clusters measurably improves tone mapping and…","keywords":["image clustering","image enhancement","tone mapping","denoising","dataset resampling","luminance descriptors","hue descriptors","unsupervised learning"],"falsifier":"Re-run the oversampling experiment on a tone-mapping network trained with CARDIE clusters selected by the two-parameter KS test, but on a dataset where the ground-truth operator is strongly local (e.g., local tone curves): if PSNR gains vanish or reverse, the relevance criterion is an artifact of the global curve family. Alternatively, replace the fitted curve with a per-pixel gain map and recompute the cluster-difference indicator; clusters that become indistinguishable under the richer model would show that the claimed relevance did not survive a more complete description of the operator.","tokens_in":9456,"feed_emoji":"🖼️","tokens_out":11850,"duration_ms":97818,"temperature":0.7,"pith_summary":"This paper introduces CARDIE, an unsupervised algorithm that groups images by luminosity and hue descriptors rather than by semantic content, and argues that these photometric clusters are the relevant units for image enhancement. Its evidence is a new quantification method: fitting a two-parameter luminance-response curve to each input-output image pair and testing, with Kolmogorov-Smirnov tests, whether clusters undergo statistically different transformations. On the 5K tone-mapping dataset, CARDIE clusters differ exactly where the tone-mapping network behaves differently, while semantic clusters from scene recognition and manual labels are nearly indistinguishable under the same test. Oversampling CARDIE's minority clusters improves tone-mapping PSNR by 0.22 dB and denoising PSNR by 1.18 dB on full test sets, with minority-class gains of 0.60 dB and 2.73 dB. If correct, the paper shows that a cheap, label-free color-and-luminosity rebalancing of training data can beat semantic clustering and random oversampling for image enhancement.","feed_headline":"Clusters by light and color beat semantic classes for image enhancement","feed_subtitle":"Oversampling those clusters adds 0.22 dB PSNR for tone mapping and 1.18 dB for denoising.","key_machinery":"The central object is a boolean descriptor per image: one luminosity flag (low, average, or high, from the median against the 20th and 80th percentiles of a $K$-image average luminance) plus a set of dominant hue bins, each flagged when at least $1/l$ of the hue distribution falls in a $\\Delta\\theta$ window around a reference color. These descriptors reduce an image dataset to a small tabular dataset, which HDBSCAN clusters with the Jaccard distance; the minimum-cluster-size and variance-threshold hyperparameters are selected via silhouette score. The companion machinery is the quantification method: fit the two-parameter Naka-Rushton curve $L_{\\mathrm{pp}} = L_{\\mathrm{i}}^{\\gamma}/(L_{\\mathrm{i}}^{\\gamma}+\\mu^{\\gamma})$ to 100 sampled pixels per image pair, then run two-sided Kolmogorov-Smirnov tests on the fitted $\\gamma$ and $\\mu$ distributions, and on local-variance changes $\\Delta\\bar{\\sigma}$ for denoising, to form the cluster-difference indicator $I^A_{l,a}$. This pair of tools converts \"are these clusters relevant?\" into a concrete statistical comparison and makes the descriptors cheap enough to compute on any enhancement dataset.","core_discovery":"On its own terms, the paper discovers that the clusters that matter for image enhancement are defined by low-level photometric properties, not by objects or scenes. CARDIE computes a per-pixel average luminance over a subsample of the dataset, flags each image as dark, average, or bright by comparing its median luminance with the 20th and 80th percentiles of that average, and records which of five hue windows contain enough of the hue distribution; the resulting boolean vectors are clustered with HDBSCAN under the Jaccard distance. The relevance test fits $L_{\\mathrm{pp}}=L_{\\mathrm{i}}^{\\gamma}/(L_{\\mathrm{i}}^{\\gamma}+\\mu^{\\gamma})$ to each input-output pair, obtains distributions of $\\gamma$ and $\\mu$ per cluster, and uses two-sided KS tests to build the cluster-difference indicator. The paper reports that CARDIE's clusters are statistically distinct exactly where the tone-mapping network applies different transformations—by luminosity level—whereas semantic clusters are not. It then shows that oversampling the minority CARDIE clusters three times improves tone mapping by 0.22 dB PSNR and denoising by 1.18 dB PSNR on the full test sets, with minority-class improvements of 0.60 dB and 2.73 dB, outperforming random oversampling and semantic-cluster oversampling.","pith_inferences":["Editorial extension: for operators with strong local or content-dependent effects, such as dehazing or local tone mapping, the two-parameter curve used in the relevance test may be too rigid; a richer per-pixel gain model could reveal that CARDIE's clusters are not the relevant ones for those operators.","A natural next experiment, not run in the paper, is to make the hue bins and luminosity thresholds adaptive per dataset and check whether the PSNR gains scale; the paper fixes these by stability rather than optimizing them.","Because resampling only changes training-data composition, combining CARDIE oversampling with architectural or augmentation improvements could compound the reported gains, and the same descriptors could be used to detect distribution shift in incoming image streams."],"forward_implications":["A practical corollary is that image-enhancement training sets can be audited and rebalanced without any labels or pretrained semantic networks; the paper shows this improves PSNR for both tone mapping and denoising.","The cluster-difference indicator can be used on any input-output enhancement dataset to find which splits of the data correspond to genuinely different operator behavior, before any training happens.","Because the descriptors are luminosity and hue only, the approach transfers to any dataset where enhancement effects are global and photometric, and the additional raw-pipeline results indicate it also works for classical processing pipelines.","The resampling procedure is architecture-agnostic: it changes only the training set, so it can be combined with any tone-mapping or denoising network, not only the two used in the paper."],"supporting_citations":[{"why":"provides the two-parameter luminance-response model used to quantify how tone mapping changes each image.","marker":"[16]"},{"why":"supplies HDBSCAN, the density-based clustering algorithm that turns the boolean descriptors into clusters.","marker":"[13]"},{"why":"provides the 5,000 input-expert-output image pairs that form the main tone-mapping dataset.","marker":"[19]"},{"why":"defines the tone-mapping network whose training is resampled and evaluated.","marker":"[18]"},{"why":"defines the denoising network whose training is resampled and evaluated.","marker":"[1]"},{"why":"supplies the Poisson-Gaussian noise model used to synthesize the denoising dataset.","marker":"[20]"},{"why":"provides the semantic scene-recognition attributes that serve as the comparison clustering baseline.","marker":"[6]"},{"why":"provides the silhouette score used to select HDBSCAN hyperparameters.","marker":"[15]"},{"why":"provides the additional raw-pipeline image pairs used to test cluster relevance outside learned operators.","marker":"[21]"}],"fun_headline_variants":["Light and color clusters beat semantic ones for enhancement","Oversampling photometric clusters boosts PSNR in tone mapping","CARDIE: clusters by luminance improve enhancement performance","Photometric descriptors yield better clusters for image enhancement","Luminance-based clustering ups PSNR for tone and denoise tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on a single premise: that the two-parameter luminance-response curve $L_{\\mathrm{pp}}=L_{\\mathrm{i}}^{\\gamma}/(L_{\\mathrm{i}}^{\\gamma}+\\mu^{\\gamma})$ summarizes what a tone-mapping operator does to an image, so that statistically different fitted parameters between clusters indicate genuinely different treatment; local, content-dependent, or clipped adjustments outside this curve family would break that link.","fun_headline_variants_meta":{"raw":{"variants":["Light and color clusters beat semantic ones for enhancement","Oversampling photometric clusters boosts PSNR in tone mapping","CARDIE: clusters by luminance improve enhancement performance","Photometric descriptors yield better clusters for image enhancement","Luminance-based clustering ups PSNR for tone and denoise tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1502,"prompt_tokens":956,"completion_tokens":546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":572,"tokens_out":546,"duration_ms":4631,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:18:22.532853+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the oversampling experiment on a tone-mapping network trained with CARDIE clusters selected by the two-parameter KS test, but on a dataset where the ground-truth operator is strongly local (e.g., local tone curves): if PSNR gains vanish or reverse, the relevance criterion is an artifact of the global curve family. Alternatively, replace the fitted curve with a per-pixel gain map and recompute the cluster-difference indicator; clusters that become indistinguishable under the richer model would show that the claimed relevance did not survive a more complete description of the operator.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the silhouette score used to select HDBSCAN hyperparameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the two-parameter luminance-response model used to quantify how tone mapping changes each image."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies HDBSCAN, the density-based clustering algorithm that turns the boolean descriptors into clusters."},{"cited_title":"Bychkovsky, S","cited_arxiv_id":null,"evidence_quote":"provides the 5,000 input-expert-output image pairs that form the main tone-mapping dataset."},{"cited_title":"Gharbi, J","cited_arxiv_id":null,"evidence_quote":"defines the tone-mapping network whose training is resampled and evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the denoising network whose training is resampled and evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Poisson-Gaussian noise model used to synthesize the denoising dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the semantic scene-recognition attributes that serve as the comparison clustering baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the additional raw-pipeline image pairs used to test cluster relevance outside learned operators."}],"review_version":1}