{"id":"0b3908e4-5f79-45b0-ab01-8bfd4b76b932","arxiv_id":"2508.08928","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper proposes a hand-built scene complexity score for light field displays and a fitted sigmoid that maps that score to the preferred depth-of-field blur, validated on three new scenes.","lead":"A new metric, DASC, scores 3D scenes by geometric detail and position relative to a light field display's depth of field. The authors fit a curve from this score to the blur level viewers prefer, and test it on a small set of new scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DASC aggregation in Eq. (20) is not a well-defined function: its branch conditions misclassify objects in front of the DoF and leave an undefined weight for inside-DoF objects.","rationale":"The reader's weakest assumption correctly identifies the DASC aggregation as the load-bearing weakness; my read agrees and sharpens it. The branch conditions in Eq. (20) do not partition the depth axis: the first branch is logically impossible, and the fourth branch swallows all objects with Zmax ≤ dΦ/2 regardless of whether they are behind, inside, or in front of the DoF. This gives negative complexity contributions to objects that are outside the DoF on the near side, directly contradicting the stated goal that outside-DoF content should increase DASC. The undefined weight for fully inside-DoF objects and the false claimed range [−1, ∞) are additional manifestations of the same defect. This is prior to the sigmoid-coefficient issue in Eq. (26), which is also real: at f = 0 the stated formula gives r ≈ 21.9, not r = 9. But without a valid f, the preference mapping cannot be meaningfully fit or validated. Therefore the verdict remains rejection; no change from the reader's assessment.","tokens_in":14724,"tokens_out":12218,"duration_ms":141221,"concrete_test":"Run the released DASC code on the nine scenes from Section V with the display parameters used in the paper, logging per-object (Zmin, Zmax, ω, ψ) and the contribution to f. Then check whether any object satisfies Zmax ≤ dΦ/2 but Zmin < −dΦ/2 (front-outside or front-spanning); if such objects receive negative contributions, the f values plotted in Figures 8 and 9 are incorrect. Also evaluate Eq. (20) literally on a synthetic scene with one object at z ∈ [−2dΦ, −dΦ] and one object fully inside the DoF: the former should yield a positive aliasing weight but instead yields −lν, and the latter triggers division by zero unless the cancellation is explicitly implemented.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires DASC value f from Eq. (18) to be computable for any scene, but the piecewise definition of ψ in Eq. (20) is not a partition of the depth axis. The first branch requires Zmin > dΦ/2 while Zmax ≤ dΦ/2, which is impossible because Zmin ≤ Zmax. More damagingly, the fourth branch is selected whenever Zmax ≤ dΦ/2, with no lower bound on Zmin. An object entirely in front of the DoF, e.g. z ∈ [−2dΦ, −dΦ], or an object spanning from the front outside region into the DoF, therefore receives ψ = −l(Ωi)/(1−ω(Ωi)) and contributes (1−ω)νψ = −lν to f in Eq. (18). Such objects are outside the DoF and alias, so they should increase scene complexity; instead they decrease f, unlike objects outside behind the DoF, which receive the positive ζ weight. The bullet in Section IV-C also defines the fully-inside-DoF weight as −l/(1−ω), which is undefined at ω = 1 unless one silently cancels the (1−ω) factor from Eq. (18); even with cancellation, front-outside objects with depth span larger than dΦ produce f values below the claimed lower bound of −1. Since f is the sole input to the preference model, the reported correlation and validation results cannot be reproduced or generalized until the piecewise definition is corrected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DASC, a depth-of-field-aware scene complexity metric that combines per-object geometric factors (entropy, edge density, curvature standard deviation) and positional factors (DoF overlap, depth position, depth span) into a single score f. The score is then mapped through a fitted sigmoid to predict the aperture radius that viewers prefer on a light field display. The authors describe a paired-comparison subjective study on a Holovizio 722 RC display with six correlation-study scenes and three validation-study scenes, use the Bradley-Terry model to derive opinion scores, fit the sigmoid, and report that the validation scenes align with the model predictions.","tokens_in":15078,"tokens_out":8523,"duration_ms":89505,"significance":"If the metric were well-defined and the model reproducible, the work would be practically useful for content creators and would address a real gap: prior light field quality assessment is mostly done on 2D displays, whereas this study evaluates DoF-guided rendering on an actual light field display. The paper also contributes a custom synthetic light field dataset and a subjective testing protocol, and states that code and data are public. However, the central mathematical definition of DASC is internally inconsistent in ways that affect every computed scene score, and the reported sigmoid coefficients contradict the accompanying textual claim. These issues must be corrected before the empirical results can be trusted.","major_comments":[{"comment":"The first branch of the piecewise definition of ψ is impossible: it requires Zmin(Ωi) > dΦ/2 and Zmax(Ωi) ≤ dΦ/2, which cannot hold because Zmin ≤ Zmax. Consequently, the piecewise function does not partition the depth axis and no case is defined for objects entirely in front of the DoF. Any object with Zmax ≤ dΦ/2, including a front-outside object at z ∈ [−2dΦ, −dΦ], falls into the fourth branch and receives ψ = −l/(1−ω); after multiplication by (1−ω) in Eq. (18), this contributes −lν to f, decreasing the complexity score for an object that is outside the DoF and prone to aliasing. This contradicts the stated modeling assumption and affects all subsequently reported DASC values.","section":"Eq. (20)"},{"comment":"The bullet defining the weight for objects fully inside the DoF as −l(Ωi)/(1−ω(Ωi)) is not a well-defined function because for a fully inside-DoF object ω(Ωi) = 1, making the denominator zero. Even if one silently cancels the (1−ω) factor against the same factor in Eq. (18), the resulting contribution is −l(Ωi)ν(Ωi), which can push f below the claimed lower bound of −1 whenever an object's depth span exceeds dΦ. The definition must be rewritten as a total function, either with explicit cancellation or with a separate branch for ω = 1.","section":"Section IV-C, bullet list and Eq. (18)"},{"comment":"The reported sigmoid coefficients κ = 21.9, β = 4.5, γ = 9.0 are inconsistent with the accompanying statement that scenes with f between −1 and 0 receive r = 9. With these coefficients, r(−1) = 21.9/(1+e^{4.5(−10)}) ≈ 21.9 and r(0) ≈ 21.9, not 9; the value r = 9 occurs near f ≈ 9.1. As written, Eq. (26) and the stated coefficients cannot reproduce the mapping that the paper claims, so the preference model is not reproducible and the plotted correlation in Figure 8 cannot be checked.","section":"Section VI-B, Eq. (26)"},{"comment":"The preference model is fitted using only six scenes that exhibit essentially two distinct preferred radii (r = 0 and r = 9). A three-parameter sigmoid fitted to two output levels is not identifiable, and the reported κ, β, γ cannot be meaningfully estimated from such data. The three-scene validation study is too small to support the strong conclusion that the model 'efficiently predicts the preferred level of blurring for any given 3D scene'; the authors partly acknowledge this in Section VII, but the generalization claim should be substantially tempered.","section":"Section VI-B and VI-C"}],"minor_comments":[{"comment":"There is a typo: 'max s)' should be 'max(s)'.","section":"Eq. (12)"},{"comment":"The depth notation is inconsistent: Eqs. (15)-(16) use Z(Ωi), Eq. (17) writes 'max(ZΩi)' with a missing parenthesis, and Eq. (20) uses Zmin/Zmax without mentioning Ωi. Please unify the notation.","section":"Eqs. (15)-(17) and (20)"},{"comment":"The text refers to 'the optimal radius r̂ mentioned in Table I', but Table I contains only likelihood-ratio values and no r̂ value; please clarify where r̂ is defined.","section":"Section VI-B"},{"comment":"The DASC values f for the individual scenes are not reported in a table, so readers cannot verify the sigmoid fit or the claimed correlations. A table with f, preferred radius, and opinion-score peaks for each scene would greatly improve reproducibility.","section":"Figure 8 and Section VI"},{"comment":"The sentence 'the score value f is scaled by the number of object m' should read 'number of objects m'.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript currently contains load-bearing mathematical errors: Eq. (20) has an impossible branch and a missing case for front-outside objects, the fully-inside-DoF weight is undefined, and the reported sigmoid coefficients cannot produce the stated f-to-r mapping. These are fixable in principle, but the corrected metric may change all DASC values and require re-fitting the model. I would not support acceptance unless the authors provide corrected definitions, recomputed scene scores, and a reproducible fit. The dataset and display-based subjective protocol are assets, but the current central claim is not supported as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: genuinely useful idea — a scene complexity metric that accounts for a light field display's depth-of-field to predict preferred DoF blur — and real work behind it: nine full-parallax densely-sampled light fields, a paired-comparison subjective study on an actual Holovizio display, and public code/data. The dataset and protocol are a solid empirical contribution.\n\nThe problem is that the central metric, DASC, is not well-defined as written. In Eq. (20), the first branch requires Zmin > dΦ/2 while Zmax ≤ dΦ/2, which cannot happen because Zmin ≤ Zmax. More damaging, the fourth branch lumps objects fully inside the DoF with objects fully in front of it. For fully inside objects ω=1, so the weight −l/(1−ω) is undefined; even if you cancel the (1−ω) factor in Eq. (18), a front-outside object gets a negative contribution to f, which is the wrong sign for an aliasing-prone region. This can also push f below the claimed lower bound of −1 for objects spanning more than dΦ. So the metric is not computable or meaningful for common scene configurations.\n\nThe sigmoid model also contradicts its own coefficients. They report κ=21.9, β=4.5, γ=9.0 in Eq. (26) and say f in [−1,0] yields r=9. But plugging those values gives r≈22 for that range. Something is off in the equation, the coefficients, or the description.\n\nValidation is thin: three new scenes, no error bars or statistical comparison, and a three-parameter sigmoid fit to six scenes with only two distinct outputs (r=0 and r=9). That's essentially a threshold on f, not a general predictive model. The authors do list limitations (small sample, few scenes, missing extremes), which is honest, but the closing claim that the model 'efficiently predicts' the preferred blur for any given scene is overreach.\n\nBottom line: the concept is worth pursuing, the dataset/protocol have value, and the flaws are fixable in a major revision. In current form the central claims don't hold. I'd send it to peer review — a good referee would catch these issues immediately — and then likely reject or require heavy revision.","headline":"Genuine empirical effort, but the DASC metric is ill-defined and the validation is too thin to support the predictive claim.","tokens_in":15574,"tokens_out":7975,"would_cite":false,"duration_ms":74915,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single scene-complexity score predicts viewers' preferred blur level on light field displays.","keywords":["light field display","depth of field","3D visualization","scene complexity metric","aliasing artifacts","DoF rendering","subjective quality assessment","perceptual preference"],"falsifier":"Take a scene containing one object entirely inside the display's DoF, so ω(Ω)=1, and compute Eq. (18) using the fourth case of Eq. (20): the weight becomes −l(Ω)/(1−ω(Ω)), a division by zero, so f is undefined. A reader could also check whether the first case of Eq. (20) is ever satisfiable for a real object; if not, the aggregation rule needs correction before the sigmoid model can be evaluated.","tokens_in":14513,"feed_emoji":"🖥️","tokens_out":6702,"duration_ms":71124,"temperature":0.7,"pith_summary":"This paper proposes a scene-complexity metric, DASC, that turns a 3D scene's geometry, texture, and position relative to a light field display's depth-of-field (DoF) into a single number. It then fits a sigmoid curve to subjective preference data so that the DASC value directly gives the blurring level—the circular-filter aperture radius—that viewers prefer when the scene is shown on such a display. If the model holds, content creators could skip subjective testing and set the DoF-rendering blur automatically. The claim is supported by paired-comparison experiments on nine synthetic light fields and a validation study with new scenes and participants.","feed_headline":"Scene complexity score predicts viewers' preferred blur level","feed_subtitle":"A fitted sigmoid turns depth, texture, and object placement into the aperture radius chosen in subjective tests.","key_machinery":"The load-bearing machinery is the DASC aggregation f=(1/m)Σ_i(1−ω(Ω_i))ν(Ω_i)ψ(Ω_i). Here ν is the mean of normalized curvature standard deviation, edge density, and entropy; ω is the fraction of the object's pixels inside the display's DoF; and ψ is a piecewise weight depending on where the object's depth range sits relative to the DoF. This condenses each object's geometric complexity and positional risk into one score. The preference mapping is then a sigmoid, r(f)=κ/(1+$e^{{β(f−γ)}}$), which is the device that converts the metric into a concrete rendering parameter.","core_discovery":"The central claim is that a single scalar, f, computed from per-object entropy, edge density, curvature variation, size, and placement relative to the display's DoF, predicts the preferred aperture radius r for DoF rendering on a light field display. The paper models the mapping as a sigmoid r(f)=κ/(1+$e^{{β(f−γ)}}$) with fitted coefficients κ=21.9, β=4.5, and γ=9.0, calibrated against Bradley-Terry preference scores from the correlation study. Scenes that sit essentially inside the display's DoF receive a mapping that avoids harmful blurring, while more complex scenes with content outside the DoF map to stronger filtering. A validation study using different scenes and a fresh set of participants reproduced the predicted preferred radii, so the paper claims the metric transfers across content.","pith_inferences":["The fitted coefficients κ, β, and γ are tied to the specific display's angular sampling and to the filter-radius units; recalibrating them for other light field displays is a direct test of the model's generality, which the paper itself lists as future work.","Because f depends only on geometry and texture, the same DASC computation could be applied to real captured light fields, not just synthetic renders, as a test of whether the preference curve survives realistic content; the paper's dataset is entirely synthetic.","The two-peak preference structure could also be modelled per observer rather than per scene: if a minority of viewers do have consistent preferences across scenes, a personalized version of the sigmoid might improve predictions for those viewers."],"forward_implications":["For a new scene, a content creator can render segmentation and depth maps, compute f, and read the preferred aperture radius directly from the sigmoid, removing the need for per-scene subjective tests.","DoF rendering can be automated: the same pipeline that detects aliasing-prone regions can set the circular-filter radius from f, keeping in-DoF content sharp and blurring only where needed.","The model gives a concrete target for content preparation: alias-free scenes should be shown without additional blur, while complex scenes with substantial out-of-DoF content should receive stronger filtering.","Subjective preference appears content-dependent rather than purely observer-dependent, so a metric based on scene properties is the right type of predictor; only about a fifth of observers in the study had preferences stable across scenes."],"supporting_citations":[{"why":"Supplies the circular-filter DoF rendering method used to generate the blur levels r applied to each light field.","marker":"[3]"},{"why":"Defines the display system's angular and spatial resolution behavior that sets the DoF region, and provides the display used in the experiments.","marker":"[4]"},{"why":"Provides the quantitative relation between pixel size, angular sampling, and distance that defines p(z) and thus the display's DoF.","marker":"[7]"},{"why":"Supplies the full-parallax densely sampled light field representation and the angular sampling rate αc needed for rendering.","marker":"[2]"},{"why":"Provides the double-stimulus impairment scale protocol that structures the paired-comparison subjective tests.","marker":"[38]"},{"why":"Provides the Bradley-Terry model used to convert paired preferences into opinion scores for each aperture radius.","marker":"[40]"},{"why":"Supplies the chi-square likelihood-ratio test used to decide whether two aperture radii are equally preferred.","marker":"[42]"}],"fun_headline_variants":["One metric to set blur for light field 3D displays","Scene complexity predicts ideal DoF blur on light field screens","New scoring model tunes blur on light field displays","DASC metric maps scene complexity to preferred blur","Predicting viewer blur preference with a scene complexity score"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the piecewise aggregation in Eq. (20) is well-defined for real scenes: in particular, the case that weights an object fully inside the DoF by −l/(1−ω) is only meaningful when ω≠1, and the listed first case (Zmin>dΦ/2 with Zmax≤dΦ/2) cannot actually occur. If those clauses are typos, the intended correction must still make f computable for every scene the model is applied to; if they are not, DASC is undefined exactly for the scenes used to fit and validate the curve.","fun_headline_variants_meta":{"raw":{"variants":["One metric to set blur for light field 3D displays","Scene complexity predicts ideal DoF blur on light field screens","New scoring model tunes blur on light field displays","DASC metric maps scene complexity to preferred blur","Predicting viewer blur preference with a scene complexity score"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1566,"prompt_tokens":944,"completion_tokens":622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":560,"tokens_out":622,"duration_ms":6283,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:33:16.464796+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene containing one object entirely inside the display's DoF, so ω(Ω)=1, and compute Eq. (18) using the fourth case of Eq. (20): the weight becomes −l(Ω)/(1−ω(Ω)), a division by zero, so f is undefined. A reader could also check whether the first case of Eq. (20) is ever satisfiable for a real object; if not, the aggregation rule needs correction before the sigmoid model can be evaluated.","supporting_citations":[],"review_version":1}