{"id":"112c4636-0f1c-47f1-ad97-7c4498970010","arxiv_id":"1908.06557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"After any tone-mapping operator, replace the maximally saturated color (the hue anchor) of each output pixel with that of the original HDR pixel, reducing measured hue distortion.","lead":"This paper proposes a post-processing step for tone mapping that shifts each tone-mapped pixel onto the constant-hue plane of the original high-dynamic-range image. The reported measurements show smaller hue errors than standard tone mappers and than a prior color-correction formula, with roughly unchanged image-quality scores.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15) forces RGB ratios to match, but CIEDE2000 hue is not guaranteed to improve; Table 2 even shows 'rosette' getting worse for Fattal and Shan, so the 'any TMO' claim is unsupported.","rationale":"The paper has a clean gamut-safety argument: since a'_w, a'_c, a'_k are nonnegative and sum to one, and c_H lies in [0,1]^3, Eq. (15) outputs a convex combination inside the RGB cube, so clipping is avoided. I agree with the reader that the central risk is the assumed equivalence between the RGB constant-hue plane and perceived hue. My read strengthens this concern with evidence internal to the paper: Table 2 shows per-image regressions for 'rosette' under Fattal and Shan, so the universal 'any TMO' conclusion is contradicted by the authors' own data. The correct verdict remains conditional: the method may work on average and is worth publishing after the claim is narrowed and the failing cases are explained. I do not see a reason to reject the paper outright, because the underlying geometric correction is well-defined and the majority of reported cases improve. The reader's conditional verdict is therefore unchanged, with the caveat made sharper.","tokens_in":10412,"tokens_out":5675,"duration_ms":59304,"concrete_test":"Re-run the §4.2 evaluation per image and per TMO, before any mean-luminance adjustment, and list paired ΔH values for conventional versus proposed for all 10 images × 6 TMOs. Check specifically whether 'rosette'/Fattal and 'rosette'/Shan reproduce the Table 2 regressions (ΔH increasing while Δc decreases); if they do, the claim that the method reduces hue degradation for any TMO is false. As an additional analytical check, take a moderately desaturated HDR pixel, for example (0.2, 0.4, 0.6), and a tone-mapped or quantized version whose normalized color differs from c_H; compute CIEDE2000 ΔH before and after Eq. (15) over a grid of a'_w and a'_c values to determine whether the correction is monotone in perceptual hue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proposal's central step, Eq. (15), replaces the tone-mapped pixel's normalized RGB color c' with the original HDR pixel's normalized color c_H while keeping the convex coefficients a'_w, a'_k, a'_c unchanged. This guarantees matching in the geometric constant-hue plane defined by Eq. (8), but not in the perceptual hue measured by CIEDE2000: CIEDE2000 hue is computed from CIELAB a*, b*, which are nonlinear functions of the RGB components and depend on the retained coefficients and on the mean luminance level. Therefore, the paper's claim that hue degradation is reduced for any TMO does not follow from the construction. The paper's own Table 2 contains concrete counterexamples: for image 'rosette', the proposed method increases the CIEDE2000 hue difference under Fattal (0.8894 to 0.9286) and under Shan (0.8983 to 0.9128), while the geometry-based Δc metric improves. The average over images hides these per-pair regressions. Near-gray pixels make the issue worse because Eq. (8) divides by max-min, so c_H is unstable or undefined when the HDR pixel is nearly achromatic. At minimum, the conclusion should be restricted to typical or average behavior and should disclose the failing cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-processing hue compensation scheme for tone-mapped images. After an arbitrary conventional tone mapping operator (TMO) produces a low dynamic range (LDR) image, each pixel is expressed in the RGB constant-hue plane as a weighted combination of white, black, and the pixel's maximally saturated color (Equations (8)-(10)). The compensation step (Equation (15)) replaces the tone-mapped pixel's maximally saturated color c' with the maximally saturated color c_H computed from the original high dynamic range (HDR) pixel, while keeping the coefficients unchanged. This places the compensated pixel on the same geometric constant-hue plane as the original HDR pixel and, because c_H is in [0,1]^3, avoids gamut clipping. The method is evaluated on 10 HDR images with six TMOs using three metrics: the geometric color difference Δc, the CIEDE2000 hue difference ΔH, and TMQI. The paper reports average improvements in Δc and ΔH, similar TMQI scores, and a favorable comparison with Mantiuk's color correction formula.","tokens_in":10694,"tokens_out":7208,"duration_ms":67340,"significance":"The proposed method is simple, parameter-free, applicable as a post-processing step to any TMO, and correctly avoids gamut clipping; these are genuine strengths, and the derivation of the compensation formula in Section 3.2 is straightforward and clearly presented. The per-image results in Tables 1 and 2 provide useful empirical information. However, the central claim that the method reduces hue degradation for all TMOs is not supported by the paper's own data, and the construction guarantees only a geometric hue match, not a perceptual one. With appropriate scoping of the claims and an explicit treatment of achromatic pixels, this could be a useful contribution; as written, the overstatement undermines the conclusion.","major_comments":[{"comment":"The claim in Section 4.2 that \"the proposed method outperformed the conventional approach for all TMOs\" is contradicted by the CIEDE2000 hue differences for image 'rosette': with Fattal's TMO, ΔH increases from 0.8894 to 0.9286, and with Shan's TMO, it increases from 0.8983 to 0.9128. These are not negligible differences and they concern exactly the pairwise comparison the abstract and conclusion rely on. The conclusion should be restricted to the observed average behavior over the test set, and these failure cases should be explicitly disclosed and, if possible, analyzed.","section":"Table 2, Section 4.2"},{"comment":"Equation (15) ensures that the normalized RGB vector of the compensated pixel equals c_H, but this geometric equality in RGB space does not imply a reduction in CIEDE2000 hue difference ΔH, because ΔH is computed from nonlinear transformations to CIELAB and depends on the retained coefficients a'_w, a'_k and on the display luminance. The empirical average improvements in Table 2 are therefore not a proof of the statement that the method reduces hue degradation \"for any TMO\"; the two worsening cells in Table 2 confirm that the guarantee is false as stated. The paper should frame the hue-preservation claim as an empirical average improvement rather than a mathematical consequence of the construction.","section":"Section 3.3, Equation (15)"},{"comment":"The Δc metric measures the Euclidean distance between the maximally saturated colors c1(i) and c2(i), which is exactly the quantity the proposed operation replaces (c' with c_H). Consequently, the large improvement in Table 1 is substantially by construction and should not be presented as independent evidence for hue preservation. The independent evidence must come from the CIEDE2000 hue difference ΔH and, secondarily, from TMQI; the paper should say so explicitly so that readers do not interpret Table 1 as validating the perceptual benefit.","section":"Section 4.2, definition of Δc"},{"comment":"Equation (8) defines the maximally saturated color using division by max(x)-min(x), which is undefined when max(x)=min(x), i.e., for achromatic pixels. The paper does not specify how the proposed method handles such pixels, either in the original HDR image or in the tone-mapped LDR image. Natural images contain many gray and near-gray pixels, so this is a real gap in the procedure. A fallback should be defined (for example, leaving achromatic pixels unchanged) and the effect of that fallback on the reported metrics should be evaluated.","section":"Section 2.2, Equation (8)"}],"minor_comments":[{"comment":"The text says \"using four TMOs\" but then enumerates six TMOs (Reinhard global, Reinhard local, Drago, Fattal, Shan, Gu). The count should be corrected.","section":"Section 4.1"},{"comment":"The phrase \"the our method\" should be corrected to \"our method\" or \"the proposed method.\"","section":"Section 4.3"},{"comment":"The paper uses the term \"constant-hue plane\" to describe a geometric construction in RGB space, but perceptual hue is measured by CIEDE2000. The paper should explicitly state that the RGB constant-hue plane is a geometric proxy for hue and that the perceptual evaluation is carried out separately with ΔH.","section":"Sections 2.2 and 3.2"},{"comment":"The caption reports ΔH=2.5178 for (b) and ΔH=2.9721 for (c), while the only difference between (b) and (c) is a mean-luminance adjustment. Since global luminance scaling should not change hue, the reason for the ΔH change should be explained (e.g., the effect of rounding quantization on the compensated values).","section":"Figure 8"}],"recommendation":"major_revision","confidential_remarks":"The core idea is a simple, potentially useful post-processing step for tone-mapped images, and the derivation is sound. The main obstacles to acceptance are the overstatement of the universal hue-preservation claim (contradicted by two cells in Table 2) and the missing treatment of achromatic pixels. Both are fixable with a moderate revision. The comparison with Mantiuk's method uses only three images and could be strengthened, but that is not blocking. The paper fits the scope of IEICE Transactions on Fundamentals, provided the claims are scoped to observed average behavior."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest but legitimate idea, presented with cleaner math than the claims. The method — post-hoc swap the tone-mapped pixel's normalized color c' for the HDR pixel's maximally saturated color c_H on the constant-hue plane (Eq. 15) — is simple, gamut-safe, and independent of the underlying TMO. That is a real contribution, even though the constant-hue plane and maximally saturated colors are not new. What is new is applying them as a TMO-agnostic post-processing step and explicitly calling out clipping and rounding as hue-distortion sources. The CIEDE2000 numbers mostly support the claim, and Table 3 shows TMQI is roughly preserved. Credit where due: the compensation formula is clear, the gamut argument is sound, and the paper does not oversell the mechanics.\n\nThe soft spots are in the scope of the claims, not the mechanics. First, the paper says the proposed method outperforms conventional TMOs for \"all TMOs\" based on averages. Its own Table 2 contains counterexamples: for 'rosette', ΔH worsens under Fattal (0.8894 to 0.9286) and Shan (0.8983 to 0.9128). Those cells should force a softer conclusion like \"usually reduces average hue difference\" plus a disclosure of where it does not. Second, Δc is nearly circular: the method replaces c' with c_H, so a smaller Δc is built into the construction; the independent evidence has to come from CIEDE2000, and there it is decent but not universal. Third, Eq. (8) divides by max-min, so near-achromatic HDR pixels make c_H unstable or undefined; the paper never says what happens there. Fourth, there are no error bars or statistical tests, and the Mantiuk comparison uses only three images. No code or data is shipped.\n\nThis is not a fatal flaw. The core idea is plausible, cheap to implement, and probably useful in practice. But the \"any TMO\" claim is unsupported as written, and the near-gray case needs an explicit treatment. A serious referee should ask for restricted claims, per-image results, a small statistical analysis, and ideally code or data. I would take a look at the revised version.","headline":"A clean, modest post-processing idea that mostly works but overclaims universality; deserves serious review with revisions.","tokens_in":11216,"tokens_out":2015,"would_cite":true,"duration_ms":21381,"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":"Hue distortion in tone-mapped images can be largely removed by replacing each tone-mapped pixel's maximally saturated color with the original HDR pixel's, a post-processing step that works with any tone mapping operator.","keywords":["tone mapping","hue preservation","constant-hue plane","maximally saturated color","color distortion","clipping and rounding","HDR imaging","CIEDE2000"],"falsifier":"Take HDR images with many pixels far outside the RGB gamut and many near-gray pixels (where max minus min is small), apply the proposed compensation, and measure CIEDE2000 hue error; if hue error does not drop for out-of-gamut pixels or increases for near-gray pixels, the geometric hue proxy fails.","tokens_in":10221,"feed_emoji":"🎨","tokens_out":3727,"duration_ms":32376,"temperature":0.7,"pith_summary":"The paper argues that hue errors in tone-mapped HDR-to-LDR images come mainly from rounding and clipping pixel values, not just from the tone curve. It proposes a post-processing step: after any tone mapping operator produces an LDR image, replace each pixel's maximally saturated color (the purest color on its constant-hue plane) with the maximally saturated color of the original HDR pixel. Because the new pixel stays on the same constant-hue plane as the HDR pixel, the hue distortion is reduced without introducing clipping errors, while the luminance from the underlying tone mapper is preserved. Experiments report lower CIEDE2000 hue differences and closer maximally saturated colors for six tone mappers, with nearly unchanged TMQI quality scores.","feed_headline":"Swap the saturated color after tone mapping to keep hue","feed_subtitle":"A one-line post-filter gives LDR images the HDR original's saturated colors for any tone mapper.","key_machinery":"The constant-hue plane in RGB space, anchored by white $w=(1,1,1)$, black $k=(0,0,0)$, and the maximally saturated color $c = (x - \\min(x))/(\\max(x)-\\min(x))$. Every pixel is expressed as a linear combination $x = a_w w + a_k k + a_c c$, where the coefficients encode luminance and saturation; the compensation step in Eq. (15) splices the HDR pixel's hue into the tone-mapped pixel by replacing $c'$ with $c_H$ while keeping the tone mapper's coefficients intact.","core_discovery":"For a tone-mapped LDR pixel $x'$, expressed on its constant-hue plane as $x' = a'_w w + a'_k k + a'_c c'$, the paper's central step is to replace $c'$ with $c_H$, the maximally saturated color of the original HDR pixel computed from Eq. (8). The compensated pixel $x'' = a'_w w + a'_k k + a'_c c_H$ lies on the same constant-hue plane as the HDR pixel, and because $c_H$ has components in $[0,1]$, $x''$ remains inside the RGB gamut, so no clipping occurs. This replacement gives the tone-mapped LDR image the same maximally saturated color values as the HDR image and reduces measured hue degradation for any tone mapping operator.","pith_inferences":["If the constant-hue-plane proxy holds perceptually, the same replacement could improve gamut mapping and multi-exposure fusion, both of which struggle with hue fidelity.","Applying the substitution iteratively or in a perceptual color space could test whether the geometric hue proxy matches human color perception more closely.","The method's benefit likely depends on how far the original HDR pixel lies outside the RGB gamut; testing on extreme out-of-gamut HDR values would clarify its limits."],"forward_implications":["The scheme is a plug-in post-filter, so any existing tone mapping operator can adopt it without redesigning the tone curve.","It directly addresses hue shifts caused by rounding quantization and clipping, which earlier color-correction formulas did not consider.","Measured CIEDE2000 hue difference and maximally saturated color difference decrease for all tested tone mappers.","The small TMQI quality loss from the compensation step can be recovered by a simple mean-luminance adjustment afterward."],"supporting_citations":[{"why":"Defines the constant-hue plane and the maximally saturated color used in Eq. (8), the geometric basis of the compensation.","marker":"[15]"},{"why":"Provides the existing hue-correction baseline the proposed method is compared against and which ignores clipping and rounding effects.","marker":"[14]"},{"why":"Defines the CIEDE2000 hue-difference metric used to evaluate hue distortion.","marker":"[23]"},{"why":"Defines the TMQI index used to verify that luminance quality is maintained after compensation.","marker":"[24]"},{"why":"One of the tone mapping operators to which the proposed compensation step is applied in the experiments.","marker":"[1]"},{"why":"Supplies the set of HDR test images used in the simulation comparisons.","marker":"[25]"}],"fun_headline_variants":["One-line filter fixes hue and gamut in tone mapping","Swap saturated colors to keep hue in tone mapping","Post-filter gives LDR images HDR color without clipping","Tone map any way, then fix hue with a color swap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The RGB constant-hue plane anchored by the maximally saturated color is assumed to align with perceived hue, so swapping the saturated color pulls the tone-mapped pixel toward the hue a human would see.","fun_headline_variants_meta":{"raw":{"variants":["One-line filter fixes hue and gamut in tone mapping","Swap saturated colors to keep hue in tone mapping","Post-filter gives LDR images HDR color without clipping","Tone map any way, then fix hue with a color swap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1337,"prompt_tokens":889,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":505,"tokens_out":448,"duration_ms":4778,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:40:16.961617+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take HDR images with many pixels far outside the RGB gamut and many near-gray pixels (where max minus min is small), apply the proposed compensation, and measure CIEDE2000 hue error; if hue error does not drop for out-of-gamut pixels or increases for near-gray pixels, the geometric hue proxy fails.","supporting_citations":[{"cited_title":"Hue-preserving color contrast enhancement method with- out gamut problem by using histogram specification,","cited_arxiv_id":null,"evidence_quote":"Defines the constant-hue plane and the maximally saturated color used in Eq. (8), the geometric basis of the compensation."},{"cited_title":"Color correction for tone mapping,","cited_arxiv_id":null,"evidence_quote":"Provides the existing hue-correction baseline the proposed method is compared against and which ignores clipping and rounding effects."},{"cited_title":"The development of the cie 2000 colour-difference formula: Ciede2000,","cited_arxiv_id":null,"evidence_quote":"Defines the CIEDE2000 hue-difference metric used to evaluate hue distortion."},{"cited_title":"Objective quality assessment of tone-mapped images,","cited_arxiv_id":null,"evidence_quote":"Defines the TMQI index used to verify that luminance quality is maintained after compensation."},{"cited_title":"Photo- graphic tone reproduction for digital images,","cited_arxiv_id":null,"evidence_quote":"One of the tone mapping operators to which the proposed compensation step is applied in the experiments."},{"cited_title":"High dynamic range image examples","cited_arxiv_id":null,"evidence_quote":"Supplies the set of HDR test images used in the simulation comparisons."}],"review_version":1}