{"id":"6b631064-a9fc-4c56-93d1-b4791029e3db","arxiv_id":"1908.06239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For 360-degree images, perceived quality is driven mainly by the central 4 degrees of the viewport (fovea and parafovea), and 19 existing quality metrics fail to predict this.","lead":"This paper measures how much each retina zone, from the fovea to the far periphery, matters for perceived quality of 360-degree images, using 256 blurred-zone versions of eight VR scenes rated by 62 people. The result, that only the central 4 degrees dominate perceived quality and that 19 existing metrics fail on spatially uneven content, is relevant for foveated rendering and viewport-adaptive VR streaming.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4's zone weights are fit from only 32 MOS values per image with 9 free parameters, and the high PCC is computed on the same fitting data; the exact equal outer-zone weights signal under-identification, so the quantitative zone-impact claim is not yet established.","rationale":"I read the paper as making two claims: a qualitative one that central retinal zones dominate perceived quality, and a quantitative one that the fitted zone weights in Table 4 measure the relative impacts. The qualitative claim is plausible and consistent with prior foveated-rendering work; the negative result on 19 objective metrics is a useful contribution. The quantitative claim, however, is where the argument is least secure. The model has 9 free parameters per image fitted to 32 MOS values generated from a highly regular 8-pattern design. The fact that several outer-zone weights come out exactly equal is a red flag for non-identifiability rather than a finding of equal perceptual impact. The paper offers no confidence intervals, no cross-validation, and no release of MOS data, so the reported PCC cannot distinguish a good model from an overfit one. The reader identified a related concern in the rationale (under-identification / no validation), though the stated weakest assumption was gaze tracking; I set agreement to partial. The recommended check, leave-one-pattern-out cross-validation, would settle whether the weights generalize to unseen spatial patterns. If it fails, the qualitative central-zone conclusion can still stand, but Table 4 and the monotonicity claim should be downgraded, supporting a conditional rather than an accept verdict.","tokens_in":14579,"tokens_out":9790,"duration_ms":102024,"concrete_test":"Run leave-one-pattern-out cross-validation for each source image: fit the 9-parameter model (4 free weights + 5 logistic parameters) on 7 of the 8 quality patterns (28 stimuli) and predict the 4 MOS values of the held-out pattern; repeat for all 8 patterns. If the held-out PCC is substantially below the in-sample 0.97, or if fitted weights vary materially across folds, the Table 4 weights are overfit to the calibration set and the quantitative zone-impact claim is unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Per image, the model has five logistic parameters plus four free zone weights (sum-to-one constraint), fit to 32 MOS values formed from only 8 spatial patterns at 4 blur levels. The eight patterns are paired (P1/P5, P2/P6, P3/P7, P4/P8) so that the low-quality weight mass is either C_j or 1-C_j, where C_j are cumulative weights; this makes individual outer-zone weights weakly identified. The fitted tables show the symptom: for I3, w2=w3=w4=w5=0.024; for I5, all four outer weights are 0.087; similar equalities appear for I1, I4, and I7. These exact equalities are hallmarks of an under-identified optimum, not evidence that perifovea and periphery have equal impact. The paper's 'reliable' label is based on PCC>=0.97 and RMSE<=0.27 computed on the same data used to fit the weights and logistic; with this parameter count and structured stimulus design, high in-sample fit is expected and does not validate the parameter values. Consequently, the specific numerical claims (w1+w2>=0.737, outer weights<=0.095, monotonicity) are not established as measurements of retinal-zone importance, even though the qualitative direction may be correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how spatially non-uniform quality in omnidirectional images affects perceived quality, with the viewport divided into five retinal-zone-inspired concentric regions (fovea, parafovea, perifovea, near periphery, far periphery). A subjective experiment with 62 participants produced MOS scores for 256 stimuli (8 source images, 8 quality-variation patterns, 4 blur levels). The authors propose a zone-weighted formulation (ZWF) combined with a five-parameter logistic mapping, fit per image to the MOS data, yielding fitted zone weights w1...w5. They report that the foveal and parafoveal zones dominate perceived quality, that content characteristics modulate these weights, and that nineteen objective quality metrics, including foveal metrics, correlate poorly with the subjective scores. The paper also details the VR viewing geometry and retina region boundaries used to define the zones.","tokens_in":14743,"tokens_out":6128,"duration_ms":61083,"significance":"If the qualitative conclusion is correct, the paper provides useful evidence for foveated rendering and viewport-adaptive streaming of omnidirectional content, and the new subjective database is a resource for the community. The evaluation of nineteen objective metrics on non-uniform-quality omnidirectional content is also a contribution. However, the quantitative zone weights in Table 4 are the output of a fitting procedure with identifiability problems, and the paper's claim that the fitting is 'reliable' is based on in-sample correlation. The qualitative direction of the result is plausible and consistent with prior work, but the specific numerical claims about individual zone weights are not established by the current analysis.","major_comments":[{"comment":"The interpretation of Table 4 as retinal-zone importance depends on the assumption that each participant maintained stable fixation at the viewport center throughout the rating, as described in Section 3. However, gaze was not tracked. The authors themselves note in Section 4.2 that for images I2 and I6, participants likely looked at a large central area rather than zone Z1 alone, which contradicts the fixed-foveation premise underlying Eqs. (11)-(12) that places the foveation point at the viewport center for all stimuli. If fixation drifted toward nearby attractive objects, the eccentricity assignments and hence the fitted weights in Table 4 are not valid measurements of retinal-zone importance. This is a load-bearing assumption for the paper's central claim and should be addressed, for example by reporting eye-tracking data or by softening the retinal-zone interpretation.","section":"4.1-4.2, Tables 4-5"}],"minor_comments":[{"comment":"Equation (11) contains a typographical artifact ('vu√') before the square root; the formula should simply read d' = sqrt( ... ).","section":"Eq. (11)"},{"comment":"The description of the fovea as 'represents 5 degrees of the central visual field or an eccentricity interval between 0 degree and 2.5 degrees' is geometrically correct but could be clarified by stating explicitly that 5 degrees is the full angular diameter and the zone is [0, 2.5) degrees of eccentricity.","section":"Section 2.2"},{"comment":"The descriptions of MSE and VPSNR are identical ('calculated based on visible pixels of a viewport with equal weights'); the distinction between raw mean squared error and peak-signal-to-noise ratio should be stated explicitly.","section":"Table 6"},{"comment":"For the metrics implemented by the authors (FWQI, FWSNR, FPSNR, F-SSIM), the text says they are based on the corresponding publications but does not provide implementation details or validation against the original authors' code; a brief note on verification would improve reproducibility.","section":"Section 5.1"},{"comment":"The claims linking the variation of w1 to the attractiveness and size of central objects are post hoc interpretations without quantitative support; consider presenting them as hypotheses rather than conclusions.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's qualitative conclusion that central retinal zones dominate quality perception is plausible and consistent with prior literature, but the quantitative zone weights are overinterpreted given the identifiability limitations. I recommend major revision: the authors should be asked to re-analyze the data with an identifiability or cross-validation check, or to limit their quantitative claims to aggregate central-vs-peripheral weights. The subjective database is a useful contribution, but the analysis needs to support the stated quantitative findings. I would not recommend reject at this stage because the qualitative direction is defensible and the issues appear addressable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth sending to a referee, but the referee should not let the headline numbers through without robustness checks. What is genuinely new: a 256-stimulus database for omnidirectional images with spatially non-uniform quality, designed around five retinal zones; a per-image zone-weighted quality model; and an evaluation of 19 objective metrics on that database. The subjective experiment is more careful than most in this area (62 participants, ITU screening, random presentation, multiple sessions), and the qualitative finding—that the central 4 degrees dominate perceived quality—is plausible and consistent with earlier foveated-rendering work.\n\nThe soft spot is the quantitative zone-importance claim. For each image, the authors fit four zone weights plus five logistic parameters to 32 MOS values. The eight distortion patterns are paired (P1/P5, P2/P6, etc.) so that the outer-zone weights are weakly identified. The exact equalities in Table 4 (e.g., w2=w3=w4=w5 for I3, I5) are the symptom of that under-identification, not evidence that perifovea and periphery contribute equally. PCC above 0.97 on the same data used for fitting is expected with that many parameters; it does not validate the parameter values. So the specific numbers (w1+w2≥0.737, outer weights ≤0.095) are not yet established as measurements of retinal-zone importance. The direction is fine; the precision is not.\n\nTwo further concerns, both acknowledged in the paper: gaze was not tracked, and the participants may have looked at a larger central area than zone Z1 for some images (I2, I6); and the retinal boundaries themselves are not standardized. These are moderate, not disqualifying, but they reinforce that the quantitative weights should be treated as rough, content-dependent indices rather than universal constants.\n\nThe 19-metric evaluation is a useful negative result: even foveal metrics fail on non-uniform quality. It is a pity the four re-implemented foveal metrics and the MOS data are not released; that limits external verification. If the authors released data and implementations, this would be a much stronger contribution.\n\nRecommendation: send to peer review. The database and qualitative finding are useful, and the weakness in the quantitative model is fixable with cross-validation, confidence intervals, and ideally gaze tracking. A referee should ask for those before the numbers are used prescriptively.","headline":"This paper has a genuinely useful new database and a plausible qualitative finding, but the fitted zone weights in Table 4 are under-identified and the numerical values should not be used prescriptively.","tokens_in":15446,"tokens_out":2275,"would_cite":true,"duration_ms":20635,"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":"Fovea and parafovea dominate perceived 360-degree image quality.","keywords":["omnidirectional images","virtual reality","quality perception","retinal zones","fovea","foveated rendering","subjective quality assessment","objective quality metrics"],"falsifier":"Run the same rating experiment with eye tracking and record where participants actually fixate during each five-second judgment. If fixations frequently leave the central 2.5-degree zone, or if recomputing zone weights from measured gaze shifts substantial weight beyond 4 degrees, then the reported fovea/parafovea dominance is an artifact of the fixation assumption rather than a property of retinal-zone perception.","tokens_in":14256,"feed_emoji":"👁️","tokens_out":15391,"duration_ms":125301,"temperature":0.7,"pith_summary":"This paper sets out to establish that, for 360-degree images displayed in VR, perceptual quality is determined almost entirely by the retinal zones surrounding the foveation point—the fovea and parafovea within about 4 degrees of gaze—and that the perifovea and periphery contribute little. The authors support this with a subjective experiment on 256 non-uniform-quality stimuli from eight scenes, where each viewport is divided into five zones matching retina regions, and with a fitted zone-weighted error model. Fitted weights put the combined central 4-degree zones at 0.737 or more of total weight in every image and every zone beyond 4 degrees at 0.095 or less; the model reproduces per-image mean opinion scores with Pearson correlation at least 0.97. They also report that nineteen existing objective metrics, including foveal metrics, all fall below PCC 0.70 on the full database, so current metrics do not reflect how viewers experience spatially varying quality in VR.","feed_headline":"Fovea and parafovea dominate perceived 360-degree image quality","feed_subtitle":"New viewer ratings show quality outside the central 4 degrees can be cut sharply with little perceived loss.","key_machinery":"The load-bearing object is the zone-weighted formulation (ZWF), a simple eccentricity-weighted MSE: $\\mathrm{ZWF}=10\\log_{10}\\left(\\frac{\\mathrm{MAX}^2}{\\sum_{k=1}^{5} w_k \\mathrm{MSE}_k}\\right)$, where $\\mathrm{MSE}_k$ is the mean squared error of pixels whose eccentricity falls in zone $Z_k$ and $w_k$ is the fitted importance of that zone. The eccentricity of each pixel is computed from a VR lens model (Eqs. 1–12), so pixels are assigned to the five retinal-zone intervals rather than to arbitrary rings. Fitting the five weights and the five parameters of the logistic mapping together by least squares is what turns subjective ratings into the per-zone importance table; the high per-image correlation (PCC $\\ge 0.97$) is the evidence that the weights carry the perceptual signal.","core_discovery":"The central discovery is a quantitative map of where quality matters in an omnidirectional viewport. The authors asked 62 participants to rate stimuli whose five concentric zones—$Z_1$ $[0,2.5^\\circ)$ fovea, $Z_2$ $[2.5^\\circ,4^\\circ)$ parafovea, $Z_3$ $[4^\\circ,9^\\circ)$ perifovea, $Z_4$ $[9^\\circ,30^\\circ)$ near periphery, and $Z_5$ $[30^\\circ,\\infty)$ far periphery—were either high or low quality in eight spatial patterns and four blur levels. Fitting weights $w_k$ in the zone-weighted formulation $\\mathrm{ZWF}=10\\log_{10}\\left(\\mathrm{MAX}^2/\\sum_{k=1}^{5} w_k\\,\\mathrm{MSE}_k\\right)$, with a five-parameter logistic mapping to MOS, gave per-image weights in which $w_1$ is usually the largest (0.404 to 0.941), the fovea-plus-parafovea share $w_1+w_2$ is at least 0.737 in every image, and all weights for zones beyond 4 degrees are at most 0.095. Pearson correlation between the fitted model and MOS is at least 0.97 per image and RMSE at most 0.27. The weights also vary with content: images with a small attractive central face put almost all weight in $Z_1$, while images with a large or poorly contrasting central object distribute weight between $Z_1$ and $Z_2$. On the same database, all nineteen tested objective quality metrics achieved PCC below 0.70 after logistic mapping, so none captured this non-uniform-quality perception.","pith_inferences":["A testable extension of the paper's weighting machinery to gaze-tracked free viewing: the effective 'central region' should expand or shift toward salient objects, so fitted weights would track attended area rather than strict retinal anatomy.","The paper's own caveat implies the numeric weight table is evidence for monotone central dominance, not exact anatomical constants, since the zone boundaries at 2.5, 4, 9, and 30 degrees are not standardized.","Extrapolating to video, a streaming system could adapt per-tile quality to the current viewport center using the same per-image fitting procedure as a per-clip calibration step."],"forward_implications":["A foveated or viewport-adaptive 360-degree encoder can spend most of its bit budget on the central 4 degrees of the user's viewport and expect little perceived quality loss from blurring the periphery.","Any objective quality metric for omnidirectional content with spatially varying quality should weight pixel errors by eccentricity, and likely by content-specific attention; unweighted viewport metrics will mispredict mean opinion scores.","Perceived quality of non-uniform stimuli can be predicted with a simple weighted-MSE model rather than complex structural or foveal metrics, once per-image weights are known.","Content characteristics—particularly the size and attractiveness of the central object and the presence of nearby objects—must enter the model, since fitted central-zone weights range from 0.404 to 0.941 across scenes."],"supporting_citations":[{"why":"It is the only earlier study on omnidirectional content with non-uniform quality and supplies the retina-area division that motivates this work.","marker":"[7]"},{"why":"It reports a foveated video-coding test showing that content attention, such as a talking face, affects perceived quality, which supports the content-dependence finding.","marker":"[11]"},{"why":"It finds peripheral degradation beyond about 7.5 degrees is barely noticed and identifies the Absolute Category Rating method as best, shaping the stimulus design and rating protocol.","marker":"[12]"},{"why":"It compares eight metrics on uniformly distorted 360-degree video, giving the baseline claim that PSNR works for uniform quality but not for spatially variable quality.","marker":"[22]"},{"why":"They provide the public panorama database used for six of the eight source images, so the stimulus set depends on this content.","marker":"[30,31]"},{"why":"It supplies the standard five-parameter logistic regression used to map objective values to subjective scores, applied to both the zone-weighted formulation and the tested metrics.","marker":"[33]"},{"why":"It provides the least-squares fitting procedure for logistic parameters and weights, making the zone-weight estimates reproducible.","marker":"[34]"},{"why":"It is the peripheral-vision review noting that retinal-region boundaries are not standardized, which the paper uses as the basis for its zone intervals.","marker":"[26]"}],"fun_headline_variants":["Fovea and parafovea drive 360° image quality perception","Peripheral quality barely matters in VR image perception","Only central 4° of viewport affects perceived VR image quality","Cut quality beyond 4° without viewers noticing","VR image quality: fovea and parafovea are what count"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that participants truly fixated the viewport center during each five-second rating and that the chosen retinal zone boundaries are accurate, so the fitted weights reflect retinal-zone importance rather than where people happened to look.","fun_headline_variants_meta":{"raw":{"variants":["Fovea and parafovea drive 360° image quality perception","Peripheral quality barely matters in VR image perception","Only central 4° of viewport affects perceived VR image quality","Cut quality beyond 4° without viewers noticing","VR image quality: fovea and parafovea are what count"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1457,"prompt_tokens":1084,"completion_tokens":373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":286}},"tokens_in":700,"tokens_out":373,"duration_ms":3605,"temperature":1.0,"reasoning_tokens":286,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:53:40.864986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same rating experiment with eye tracking and record where participants actually fixate during each five-second judgment. If fixations frequently leave the central 2.5-degree zone, or if recomputing zone weights from measured gaze shifts substantial weight beyond 4 degrees, then the reported fovea/parafovea dominance is an artifact of the fixation assumption rather than a property of retinal-zone perception.","supporting_citations":[{"cited_title":"Perceptual Quality Assessment of Immersive Images Considering Peripheral Vision Impact","cited_arxiv_id":"1802.09065","evidence_quote":"It is the only earlier study on omnidirectional content with non-uniform quality and supplies the retina-area division that motivates this work."},{"cited_title":"Subjective quality evaluation of foveated video coding using audio-visual focus of attention,","cited_arxiv_id":null,"evidence_quote":"It reports a foveated video-coding test showing that content attention, such as a talking face, affects perceived quality, which supports the content-dependence finding."},{"cited_title":"Is foveated rendering perceiv- able in virtual reality?: Exploring the efficiency and consistency of quality assessment methods,","cited_arxiv_id":null,"evidence_quote":"It finds peripheral degradation beyond about 7.5 degrees is barely noticed and identifies the Absolute Category Rating method as best, shaping the stimulus design and rating protocol."},{"cited_title":"A study on quality metrics for 360videocommunications,","cited_arxiv_id":null,"evidence_quote":"It compares eight metrics on uniformly distorted 360-degree video, giving the baseline claim that PSNR works for uniform quality but not for spatially variable quality."},{"cited_title":"A statistical evaluation of recent full reference image quality assessment algorithms,","cited_arxiv_id":null,"evidence_quote":"It supplies the standard five-parameter logistic regression used to map objective values to subjective scores, applied to both the zone-weighted formulation and the tested metrics."},{"cited_title":"Q-STAR:Aperceptual video quality model considering impact of spatial, temporal, and amplitude resolutions,","cited_arxiv_id":null,"evidence_quote":"It provides the least-squares fitting procedure for logistic parameters and weights, making the zone-weight estimates reproducible."},{"cited_title":"Pe- ripheral vision and pattern recognition: A review,","cited_arxiv_id":null,"evidence_quote":"It is the peripheral-vision review noting that retinal-region boundaries are not standardized, which the paper uses as the basis for its zone intervals."}],"review_version":1}