{"id":"eaba97d0-fb13-4c99-8d60-9d53586cd1fd","arxiv_id":"2507.21557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper presents the first JND dataset for point clouds viewed in 6DoF VR, showing lower thresholds for texture than geometry distortions and a correlation between colorfulness and texture JND.","lead":"A team measured the smallest amount of compression that people can notice when viewing 3D point clouds in virtual reality. They created a public dataset of these thresholds and found that color (texture) distortions become noticeable at lower compression settings than shape (geometry) distortions for most models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that texture PCJND is smaller than geometry PCJND rests on comparing V-PCC QP values across two different distortion modalities without calibrating QP to a common perceptual or distortion scale.","rationale":"The reader's identified weakest assumption is exactly the load-bearing issue I find: the texture-vs-geometry comparison is made in V-PCC QP units, which are not calibrated across the two distortion modalities. I read the paper in good faith: it is a large, careful subjective study with screened participants, outlier removal, Grubbs tests, and SUR fitting, and the PC-JND dataset is a potentially valuable contribution if released. The QP-comparability issue is real but fixable by reframing the claim or adding a calibrated comparison, so conditional acceptance remains appropriate. Secondary issues—dataset availability, an apparent inconsistency between '51 distorted versions per reference' in the abstract and the two times 51 distortion conditions in Sec. III, and the use of a paired t-test on per-reference means obtained from different subject groups—do not change the verdict but should be addressed. Notably, the paper's own Limitation subsection covers subject bias but omits the scale-comparability threat, which is the main reason the central claim is not yet established.","tokens_in":16116,"tokens_out":8122,"duration_ms":93024,"concrete_test":"Re-evaluate the texture-vs-geometry comparison on a common distortion axis. For each of the 34 references, take the 85% PCJND QP values from Table III, encode/use the corresponding distorted point clouds, and compute (a) bitrate and (b) a point-cloud quality metric applicable to both attribute and geometry errors (e.g., PCQM, or D1-PSNR for geometry plus color PSNR on the texture map, reported side by side). Then count how many references have texture PCJND < geometry PCJND on the common metric. If the majority ordering survives, the headline is robust; if it inverts or drops near chance, the QP-scale artifact explains the result. Additionally, re-run Part II with texture mapped onto the reconstructed geometry in the reference condition to quantify how much of the 'geometry-only' JND is actually due to induced texture error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central finding (Abstract; Sec. V.b) is that texture PCJND is smaller than geometry PCJND. The evidence is a comparison of fitted V-PCC QP thresholds: mean texture PCJND = 23.70/27.20/29.99/33.71 vs geometry = 28.61/30.65/32.28/34.44 at 85/75/65/50% SUR (Table III). The load-bearing weakness is that V-PCC QP is an encoder control parameter, not a perceptual distortion scale, and the two compared conditions quantize fundamentally different signals. Texture QP acts on color attributes; geometry QP acts on projected depth/coordinate videos. Nothing in the paper calibrates these scales to a common unit (bitrate, D1/PSNR, color PSNR, PCQM, or direct perceptual anchoring), so a smaller QP threshold for texture might simply reflect that one unit of texture QP produces more visible distortion than one unit of geometry QP, not that observers are intrinsically more sensitive to texture. The manuscript even hedges in Sec. V.b: 'if QP is used to measure distortion,' but the abstract and conclusion drop the qualification. A secondary confound: in Part II, geometry QP changes the reconstructed geometry onto which color attributes are projected, so the 'geometry-only' condition still injects texture distortion; the two conditions are not cleanly isolated. The paper's own Limitation subsection (Sec. V.d) addresses response bias but not this scale-comparability threat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a subjective study of point-cloud-wise just noticeable difference (PCJND) for V-PCC-compressed point clouds viewed in a 6DoF VR environment. The authors select 34 reference point clouds, encode each with V-PCC at 51 texture-QP levels (Part I) and 51 geometry-QP levels (Part II), and collect binary difference judgments from 68 subjects using a relaxed binary search with side-by-side reference/distorted pairs in an HTC Vive environment. After outlier removal, PCJND samples are fit to normal distributions, SUR curves are derived, and PCJND thresholds are reported at 50%, 65%, 75%, and 85% SUR. The paper claims that texture PCJND is smaller than geometry PCJND for most references, that texture PCJND correlates positively with colorfulness, and that there is no clear correlation between colorfulness and geometry PCJND or between the number of points and either PCJND. The PC-JND dataset is introduced as a public benchmark.","tokens_in":16427,"tokens_out":5751,"duration_ms":69500,"significance":"If the findings hold, this is a useful first benchmark for point-cloud JND in immersive VR: the dataset of 34 references and 51 distortion levels per reference fills a clear gap, and the subjective protocol with HMD/6DoF viewing, outlier screening, and SUR analysis is broadly sound and worth publishing as a measurement contribution. The content-dependent SUR behavior and the colorfulness-texture correlation are interesting, falsifiable observations. The main weakness is that the headline comparison between texture and geometry JND is made on uncalibrated V-PCC QP scales, so the perceptual ordering does not follow from the data as presented. The dataset and measurement methodology remain valuable independent of that cross-modality comparison.","major_comments":[{"comment":"The central claim that texture PCJND is smaller than geometry PCJND (Abstract; Section V-B; Table III) compares fitted thresholds on the V-PCC QP scale across two different distortion types. QP is an encoder parameter, not an established perceptual scale, and there is no evidence that a one-step change in texture QP produces the same amount of visible distortion as a one-step change in geometry QP. The paper itself hedges in Section V-B ('if QP is used to measure distortion'), but the Abstract and Conclusion present the comparison without this qualification. Please either calibrate the two QP scales to a common unit (e.g., bitrate, D1/color PSNR, PCQM, or a direct behavioral anchor) or reframe the result as two separate JND measurements without a cross-modality ordering.","section":"Section V-B, Table III"},{"comment":"The 'geometry-only' condition in Part II may not be cleanly geometry-only. In V-PCC, geometry compression changes the reconstructed point positions onto which color attributes are mapped, so the condition with geometry QP variation and uncompressed texture video can still introduce texture/color artefacts at the rendered surface. The paper should quantify the resulting color distortion (e.g., color PSNR between the reference and the geometry-only reconstructions) or otherwise demonstrate that texture distortion is negligible, before interpreting the geometry PCJND as purely geometric sensitivity.","section":"Section III-A, Section V-B"}],"minor_comments":[{"comment":"The sentence 'the point clouds used in the test are listed in TableI' appears to reference the wrong table; the reference point clouds are listed in Table II, while Table I summarizes existing JND datasets.","section":"Section III-A"},{"comment":"There are several typos and inconsistent names: 'THC Vive' should be 'HTC Vive'; 'Unversity' should be 'University'; 'Scuptures' appears as 'Sculptures'; 'the20smaria' and 'ulliwegner' in Table II are written as 'the20sMaria' and 'UliWegner' in the text; 'A VS' is likely 'AVS'.","section":"Section III-B"},{"comment":"The PSNR columns in Table III are not clearly defined: the text mentions 'MSE, PSNR (p2point)' but the table has no MSE column, and it is not stated whether the texture PSNR is computed on color attributes and the geometry D1 on point positions. Please specify the exact metrics and their meaning at the 75% PCJND operating point.","section":"Section V-B, Table III"},{"comment":"Please clarify how the 68 subjects were assigned to the two parts and sessions; the text says 68 subjects were involved, but after outlier removal only 33 geometry and 35 texture samples per reference remain, and Table I lists 33 in the 'No. rates/seq' column. State explicitly how many valid subjects rated each reference in each part.","section":"Section III-C"},{"comment":"The correlation analysis reports coefficients but no p-values, confidence intervals, or correction for the multiple correlations tested. Please add significance tests so that the statements 'texture PCJND correlates with colorfulness' and 'there is no significant correlation' are backed by inferential statistics.","section":"Section V-C"},{"comment":"The Limitations subsection addresses response bias but does not discuss the QP-scale comparability issue raised above; an explicit acknowledgment would help readers interpret the texture-versus-geometry comparison.","section":"Section V-D"}],"recommendation":"major_revision","confidential_remarks":"This is essentially a dataset-and-measurement paper. If the authors cannot provide a convincing QP calibration, I would still consider the paper publishable provided the cross-modality ordering is removed and the two JND distributions are reported separately. Please also verify that the dataset will actually be released publicly with sufficient documentation to be usable as a benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth a serious look: it is the first JND study for point clouds in 6DoF VR, with a clean subjective protocol and a dataset that should be useful to the point cloud compression and quality assessment community. The authors ran 68 subjects through a controlled HMD experiment, used a relaxed binary search and standard outlier rejection, and fit SUR curves to derive texture and geometry JND thresholds for 34 references. The dataset (51 distorted versions per reference from V-PCC) fills a real gap, and the reported correlation between texture JND and colorfulness is a plausible new result.\n\nThe soft spot is the headline comparison. The abstract says \"texture PCJND of human eyes is smaller than the geometry PCJND,\" but the evidence is a comparison of V-PCC QP thresholds. QP is an encoder parameter, not a calibrated perceptual scale. One unit of texture QP does not equal one unit of geometry QP in visible distortion, so the finding is only meaningful operationally: at equal encoder settings, texture artifacts become visible at lower QP than geometry artifacts. The authors almost say this in Sec. V.b (\"if QP is used to measure distortion\"), but the qualification vanishes in the abstract and conclusion. That overstatement should be fixed, either by reframing or by adding a calibration (e.g., bitrate or a common distortion metric).\n\nThere is also a secondary confound in the geometry-only condition: V-PCC projects color attributes onto the reconstructed geometry, so compressing geometry can inject texture misalignment or color artifacts. The paper treats the two conditions as isolated, but they are not perfectly separable. That is worth a robustness check or at least a caveat.\n\nThe data processing looks sound: normality was checked, outliers were removed with Grubbs and subject-level criteria, and the SUR fitting is standard. The reader's worry about circularity does not land; the JND thresholds come from subjective reports rather than a model, so there is no circular reasoning.\n\nThe paper deserves peer review. The dataset alone is a contribution, and the comparison can be salvaged with clearer language. My main asks to the authors would be: make the dataset available, reframe or calibrate the texture-vs-geometry claim, and address the geometry/texture confound explicitly.\n\nBest,\n[Your name]","headline":"First point-cloud JND dataset for 6DoF VR, but the headline claim that texture JND is smaller than geometry JND overstates what an uncalibrated QP comparison can support.","tokens_in":16960,"tokens_out":3428,"would_cite":true,"duration_ms":38558,"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":"The paper reports the first point-cloud just-noticeable-difference measurements in 6DoF virtual reality: texture distortion becomes visible at a smaller V-PCC quantization step than geometry distortion for most point clouds.","keywords":["just noticeable difference","point clouds","6DoF virtual reality","V-PCC","subjective quality assessment","satisfied user ratio","colorfulness","JND dataset"],"falsifier":"Re-test the same 34 clouds with distortion levels matched on a common scale, for example equal bitrate or equal PSNR for texture and geometry, and check whether texture PCJND remains below geometry PCJND for most references; if the ordering flips or dissolves, the headline comparison is an artifact of QP units.","tokens_in":15920,"feed_emoji":"🕶️","tokens_out":12159,"duration_ms":123035,"temperature":0.7,"pith_summary":"The paper sets out to measure, for the first time, the point-cloud just noticeable difference (JND, the minimum distortion at which a difference becomes visible) in a six-degrees-of-freedom virtual-reality environment, and to publish the resulting thresholds as a reusable benchmark. Using 34 reference point clouds and a pool of 68 viewers on a head-mounted display, it finds that, with V-PCC quantization as the distortion scale, texture distortion becomes visible at a smaller quantization step than geometry distortion for most clouds. The paper also reports that the texture threshold grows with the colorfulness of the cloud, while geometry thresholds show no such link. These results give compression systems a content-dependent target: colorful clouds should spend proportionally more bits on color, and geometry and texture should be treated as separate perceptual budgets.","feed_headline":"Texture distortion becomes visible before geometry distortion in VR","feed_subtitle":"At the 85% threshold, 28 of 34 VR point clouds show lower texture than geometry JNDs; the dataset is public.","key_machinery":"The argument is carried by the PCJND protocol: side-by-side display of a reference and a distorted point cloud in a 6DoF virtual room, a relaxed binary search that narrows the QP search range by three quarters at each step, and per-reference aggregation of the resulting integer thresholds. V-PCC (Video-based Point Cloud Compression) generates the distortions by projecting each cloud into geometry and texture videos that are compressed with HEVC, and the QP of those videos is the distortion step. A satisfied-user-ratio (SUR) curve is fitted per reference as the complementary cumulative distribution function of the normal distribution fitted to the collected PCJND samples, and the PCJND is the smallest QP at which the SUR reaches a chosen threshold T (85%, 75%, 65%, or 50%). Texture PCJND and geometry PCJND come from the same protocol with only texture or only geometry compressed, so the comparison in QP units is the central instrument of the paper.","core_discovery":"Working in a 6DoF VR room where subjects could walk around and view point clouds from all sides, the study measured the point-cloud-wise just noticeable difference (PCJND) separately for texture and for geometry, using V-PCC compression levels as the step size. Distortion levels were generated by varying the texture QP from 1 to 51 while leaving geometry lossless, and vice versa. For each reference, a relaxed binary search located each subject's threshold, outlier subjects and samples were removed, and a satisfied-user-ratio (SUR) curve was fitted as the complement of a normal cumulative distribution, from which 85%, 75%, 65%, and 50% PCJND values were read. At the 85% SUR threshold, 28 of 34 references had a smaller texture PCJND than geometry PCJND, and the mean texture PCJND was smaller at every threshold; paired t-tests rejected equal means at 65%, 75%, and 85%. The texture PCJND correlates with colorfulness (r = 0.5655), whereas geometry PCJND does not, and the number of points correlates with neither. The resulting PC-JND dataset contains 34 references and 51 V-PCC-distorted versions of each, with per-reference threshold labels, and is being made publicly available.","pith_inferences":["Because the paper compares texture and geometry QP directly, an obvious extension is to recalibrate QP to a common distortion metric and re-test whether the ordering survives.","Real V-PCC encodes geometry and texture jointly, so testing mixed distortions could reveal masking between the two distortion types that this separate-distortion design cannot capture.","The colorfulness correlation suggests that a rendered-image-based colorfulness feature, rather than raw point statistics, might predict texture PCJND; this is testable on the released dataset without new viewing experiments.","The lack of correlation with point count may not generalize to sparser clouds; a synthesis experiment with matched geometry error across point densities could separate density from visible surface detail."],"forward_implications":["A point-cloud encoding pipeline can treat texture and geometry as separate perceptual budgets, allowing higher geometry QP than texture QP before visible degradation for most content.","Colorfulness gives a cheap content feature for predicting texture PCJND, potentially replacing or initializing per-content subjective tests.","The four SUR thresholds (85%, 75%, 65%, 50%) let applications choose an operating point according to how sure they need to be that no viewer notices a difference.","PC-JND provides a benchmark for training objective PCJND predictors and for validating point-cloud quality metrics under 6DoF HMD viewing."],"supporting_citations":[{"why":"Defines the V-PCC encoder whose QP settings generate the 51 texture and geometry distortion levels used in the subjective test.","marker":"[18]"},{"why":"Supplies the relaxed binary-search procedure and the outlier-subject rejection rule used to collect and clean PCJND samples.","marker":"[22]"},{"why":"Supplies the formulation of JND as the smallest QP at which the satisfied-user-ratio reaches a chosen threshold.","marker":"[21]"},{"why":"Provides the colorfulness measure whose correlation with texture PCJND is reported.","marker":"[27]"},{"why":"Supplies the outlier-sample removal and normality-testing procedure applied to the PCJND samples.","marker":"[25]"},{"why":"Provides the controlled-room and double-stimulus recommendations that the VR subjective environment follows.","marker":"[17]"},{"why":"Provides prior methodology for subjective quality assessment of compressed point clouds on a 6DoF head-mounted display.","marker":"[30]"},{"why":"Supports the choice of V-PCC over G-PCC for dense colored point clouds by comparing the codecs' compression quality.","marker":"[29]"}],"fun_headline_variants":["Texture distortion more visible than geometry in VR point clouds","Point cloud JND: texture beats geometry in 6DoF VR","VR reveals texture JND smaller than geometry for point clouds","Colorfulness predicts texture JND, not geometry, in point cloud VR","New dataset: point cloud JND thresholds for texture and geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one V-PCC quantization step on the texture scale and one on the geometry scale are perceptually comparable units, so that a smaller texture PCJND can be read as texture errors being noticed first.","fun_headline_variants_meta":{"raw":{"variants":["Texture distortion more visible than geometry in VR point clouds","Point cloud JND: texture beats geometry in 6DoF VR","VR reveals texture JND smaller than geometry for point clouds","Colorfulness predicts texture JND, not geometry, in point cloud VR","New dataset: point cloud JND thresholds for texture and geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1433,"prompt_tokens":1077,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":693,"tokens_out":356,"duration_ms":4543,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:37:05.273429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-test the same 34 clouds with distortion levels matched on a common scale, for example equal bitrate or equal PSNR for texture and geometry, and check whether texture PCJND remains below geometry PCJND for most references; if the ordering flips or dissolves, the headline comparison is an artifact of QP units.","supporting_citations":[{"cited_title":"An overview of ongoing point cloud compression standardization activities: video-based (v- pcc) and geometry-based (g-pcc),","cited_arxiv_id":null,"evidence_quote":"Defines the V-PCC encoder whose QP settings generate the 51 texture and geometry distortion levels used in the subjective test."},{"cited_title":"VideoSet: A large-scale compressed video quality dataset based on jnd measurement,","cited_arxiv_id":null,"evidence_quote":"Supplies the relaxed binary-search procedure and the outlier-subject rejection rule used to collect and clean PCJND samples."},{"cited_title":"MCL-JCV: a JND-based H. 264/A VC video quality assessment dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the formulation of JND as the smallest QP at which the satisfied-user-ratio reaches a chosen threshold."},{"cited_title":"Picture-level just noticeable difference for symmetrically and asymmetrically com- pressed stereoscopic images: Subjective quality assessment study and datasets,","cited_arxiv_id":null,"evidence_quote":"Supplies the outlier-sample removal and normality-testing procedure applied to the PCJND samples."},{"cited_title":"Methodologies for the subjective assessment of the quality of television images,","cited_arxiv_id":null,"evidence_quote":"Provides the controlled-room and double-stimulus recommendations that the VR subjective environment follows."},{"cited_title":"Subjective quality database and objective study of compressed point clouds with 6DoF head-mounted display,","cited_arxiv_id":null,"evidence_quote":"Provides prior methodology for subjective quality assessment of compressed point clouds on a 6DoF head-mounted display."},{"cited_title":"Quality evaluation of static point clouds encoded using mpeg codecs,","cited_arxiv_id":null,"evidence_quote":"Supports the choice of V-PCC over G-PCC for dense colored point clouds by comparing the codecs' compression quality."}],"review_version":1}