{"id":"b0fed612-0b29-45de-8213-2fc6e73b03a4","arxiv_id":"2506.23833","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A resolution-invariant binary image comparison metric built from distance-transform anchor points and four marked point-process summary measures.","lead":"This paper proposes a new metric, PointSSIM, for comparing binary images that is designed to be invariant to resolution and rotation. It extracts 'anchor points' from the distance transform and compares four summary statistics of the resulting point pattern.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The resolution-invariance claim is the load-bearing premise, yet the paper's only test (Section III.C) compares independent realizations across resolutions rather than PointSSIM between an image and its own rescaling, and the admitted V3/V4 deviations are never bounded.","rationale":"The paper's central contribution is a metric whose headline property is resolution invariance. That property holds only if the four summary measures are stable functions of the underlying scene, not of the pixel grid. The reader's weakest assumption pinpoints exactly this: local maxima of the distance transform, and hence anchor counts and positions, can change under rescaling. I agree with that assessment. The additional sharpening here is that the reported experiment is not a direct test: it compares ensembles of separately generated realizations at different resolutions, so the y=x scatter in Figure 18 conflates realization-to-realization variation with resolution effects. The admitted deviations in V3 and V4 are especially concerning because V3 divides by the number of connected components, which changes when low-resolution images merge objects, and V4 depends on subregion counts that are sensitive to point positions. The metric may still work well within some resolution range and object-size regime, but that regime is neither defined nor bounded. This is an addressable empirical gap, not a logical contradiction, so the conditional verdict is appropriate; the concrete test above would either validate the claim or force a precisely scoped revision. Secondary issues (absence of code/data, V3/V1 overlap, qualitative baseline comparisons) reinforce the need for the test but do not displace the invariance question.","tokens_in":11855,"tokens_out":5493,"duration_ms":63657,"concrete_test":"Use the same continuous object parameters (e.g., 50 'mixture of ellipses and circles' realizations) and render each scene at 64, 128, 256, 512, and 1024 grid resolutions; do not re-simulate per resolution. For each realization, compute PointSSIM between every pair of resolutions and also record per-measure differences V_i(high) - V_i(low). A pass criterion that would support the current claim: median across all pairs ≥ 0.95 and |mean V3/V4 difference| ≤ 0.05 over the tested resolution range. If either condition fails, the paper must specify the resolution range and object-size regime over which invariance holds, or weaken the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that PointSSIM is 'resolution invariant' requires the anchor-point representation — counts, positions, radii, and object labels from Equations 4–6 — to be approximately unchanged when the same binary scene is rendered at different resolutions. This is the condition on which the central advantage over SSIM/MSE rests. The paper's resolution experiment (Section III.C, Figures 17–18) does not establish it: the 256, 512, and 1024 images are independently generated realizations, so the scatter against the identity line measures distributional overlap, not whether a given image maps to the same descriptor after rescaling. Even in that indirect comparison, V3 (anchor points per object, Eq. 9) and V4 (spatial variance irregularity, Eq. 10) drift below the diagonal, and the text explains this by object merging and smoothing at low resolution. Section I concedes 'slight sensitivity to scale' but supplies no quantitative bound or operating regime. Until invariance is measured as PointSSIM(x, rescaled(x)) and the allowed drift is specified, a user cannot know whether a cross-resolution comparison is trustworthy; the advertised property is therefore unsupported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PointSSIM, a four-dimensional summary metric for binary images. Each image is first reduced to 'anchor points' (local maxima of the Euclidean distance transform after a local-adaptivity filter) with marks for radius and object label; the proposed metric then compares images via four scalar measures: anchor count (V1), area coverage (V2), anchor points per object (V3), and spatial variance irregularity (V4). PointSSIM is claimed to be resolution and rotation invariant, and it is benchmarked against SSIM, MSE, and MS-SSIM on five simulated image families, including a three-resolution experiment. The paper argues that the point-process compression avoids pixel-to-pixel correspondence and improves between-class discrimination.","tokens_in":12070,"tokens_out":10905,"duration_ms":118825,"significance":"If the invariance claims could be supported, PointSSIM would be a lightweight and interpretable descriptor for comparing segmented and binary images across scales, which is relevant for geostatistical simulation and structural image analysis. The algorithmic pipeline is clearly specified and the four measures are easy to compute. However, the paper's central promise - resolution invariance - rests on an unproven and partially incorrect relationship between the distance-transform anchors across resolutions, and the rotation invariance of V4 is not established. The reported dataset comparisons are suggestive but do not yet quantify the key property.","major_comments":[{"comment":"As currently defined, V2 is not invariant under resolution rescaling. The radius mark A^r in Eq. (5) is the raw distance-transform value in pixel units, and the denominator Lx*Ly in Eq. (8) is the size of the common base coordinate system, which for two differently sized images is the minimum of the two image dimensions. If the same scene is rendered at twice the resolution, A^r approximately doubles, the sum of squared radii quadruples, while Lx*Ly is unchanged, so V2 does not approximate the same value. Please convert the radii to physical units by multiplying by the cell size, or use each image's own area as the denominator, and rerun the resolution experiment with the corrected definition.","section":"Section II.A/B, Eq. (8)"},{"comment":"The resolution experiment compares 50 independent realizations generated at 256, 512, and 1024 pixels, not the same image downsampled or upsampled. The histograms and scatter around the identity line therefore measure distributional overlap between two ensembles, not whether the descriptor of a fixed scene is stable under rescaling. The text admits that V3 and V4 drift below the diagonal because of object merging at low resolution, and Section I concedes 'slight sensitivity to scale' without a bound. To support the headline property, report PointSSIM(x, rescaled(x)) and per-measure drift for fixed images across scale factors, together with a stated tolerance or operating range.","section":"Section III.C, Figures 17-18"},{"comment":"The fourth measure V4 is not rotation invariant as implemented. It partitions the domain into 10x10 axis-aligned quadratic subregions and counts anchor points per subregion; a rotation of the image relative to this fixed grid will generally change the counts and therefore V4. Since Section I claims that 'the measure comparison is rotation invariant', please test V4 on rotated versions of the same image or redefine the measure so that the subregions rotate with the image.","section":"Section II.C, Eq. (10)"},{"comment":"The claim in Section III.B that the four measures are 'not redundant' is contradicted by Eq. (9), which defines V3 = V1 / max(A_l), and by the paper's own statement in Section III.C that 'measure 3 is the same as measure 1 except that we divide by the number of objects'. V3 is a deterministic function of V1 and the object count, so weak correlation in the scatterplot of Figure 14 does not establish that the measures are independent. Please qualify this claim and consider reporting the object count separately if the descriptor is meant to be non-redundant.","section":"Section III.B and Eq. (9)"}],"minor_comments":[{"comment":"Equation (11) is garbled and ambiguous as typeset; please rewrite it with an explicit opening parenthesis and state how the metric behaves when max(Vi(x1), Vi(x2)) = 0.","section":"Equation (11)"},{"comment":"Equation (10) is used for both V4 and for s^2; please renumber the second equation.","section":"Equation (10)"},{"comment":"Several figure callouts do not match the displayed figure numbers, for example the text refers to Figure 14 when the displayed panel is labeled Figure 13, and to Figure 8 when the displayed panel is labeled Figure 7.","section":"Figure callouts"},{"comment":"The choice of 100 subregions for V4 is justified only by 'seems to work well empirically'; a short sensitivity analysis over the number of subregions would make the method more robust.","section":"Section II.C, V4 subregions"},{"comment":"No code or data are provided, so the numerical results in Figures 15-18 cannot be reproduced; including them would strengthen the paper.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know this paper is a reasonable combination of known tools packaged as a new low-dimensional image comparison metric for binary images. The advertised resolution invariance is the load-bearing claim, and the paper's own experiment does not actually test it. That makes the paper conditional at best, not a reject.\n\nWhat's new: the pipeline of distance-transform local maxima filtered by local adaptivity, then a marked point process (coordinates, radius, object label), then four summary measures combined by a normalized Euclidean distance. Each piece is standard morphology or spatial statistics, but the assembly is not in the cited literature. The measure is rotation invariant and low-dimensional, which is genuinely useful for geostatistical and remote sensing problems where binary images at different resolutions need comparing. The paper is clearly written and the synthetic experiments show nice separation between image classes.\n\nSoft spots, in order. First, the resolution invariance test in Section III.C compares independent realizations generated at 256, 512, 1024, not the same image rescaled. So it measures distributional overlap, not whether an image maps to the same descriptor after rescaling. That is the central claim, and it is unsupported. Even in that indirect test, V3 and V4 drift below the diagonal because low-resolution images merge objects; the paper acknowledges this but gives no bound or operating regime. Second, V3 is V1 divided by object count, so it is deterministically related to V1; the claim that the measures are not redundant is overstated. Third, Equation 11 is garbled as printed, and the metric has undefined cases when the max in the denominator is zero. Fourth, no code or data are provided, so the reported separation cannot be checked. The comparison against SSIM/MSE is visual, no error bars or statistical tests.\n\nWhat is good: the authors clearly state limitations, the method is efficient and interpretable, and the core idea is plausible. All weaknesses are addressable with code, a proper rescaling test, and a corrected metric definition.\n\nThis paper deserves a serious referee. It is not ready as is, but the idea is worth a revision. I would bring it to the reading group only for the discussion of what 'resolution invariance' should mean.","headline":"A sensible combination of known morphology and point-pattern tools, but the resolution-invariance claim is not actually tested and the metric's own redundancy is overlooked.","tokens_in":12592,"tokens_out":2673,"would_cite":false,"duration_ms":28851,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","62M30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new metric, PointSSIM, compares binary images of different resolutions by turning each image into a small set of anchor points and comparing four structural summaries, without any pixel-to-pixel matching.","keywords":["image comparison metric","binary images","resolution invariance","marked point process","distance transform","mathematical morphology","structural similarity","spatial point pattern"],"falsifier":"Take a binary image containing two distinct objects separated by a one-pixel gap and compare PointSSIM for the same scene rendered at 256x256 and at 1024x1024. If the gap closes at low resolution, the low-resolution version has one connected object instead of two, so V3 (anchor points per object) roughly doubles or halves, and the scatter point falls far off the identity line. A single such image pair where V1 or V3 deviates by a large margin between resolutions would falsify the resolution-invariance claim in a concrete, checkable case.","tokens_in":11629,"feed_emoji":"🖼️","tokens_out":5304,"duration_ms":50785,"temperature":0.7,"pith_summary":"PointSSIM compares two binary images without aligning their pixel grids. It first turns each image into a marked point pattern: anchor points picked as locally adaptive maxima of the distance transform, each carrying a radius and an object label. Four summary measures—anchor count, area coverage, anchor points per object, and spatial variance irregularity—are then compared between images, giving a similarity score between 0 and 1. The claim is that these measures are stable under rescaling, so images of different resolutions can be compared directly. This matters for geostatistics and other fields where structural comparisons across grid resolutions are common.","feed_headline":"Four numbers compare binary images at any resolution","feed_subtitle":"Anchor points from a distance transform keep PointSSIM stable across grid sizes and rotations.","key_machinery":"The object that carries the argument is the marked point-process representation of the image. Anchor points are the locally adaptive maxima of the minimal distance transform: grid cells whose distance to the nearest zero-valued pixel is maximal among their eight neighbors, with the extra condition that no two anchor points are closer to each other than to the edge of the object. Each anchor point carries two marks, the radius of the largest inscribed circle centered at the point and the label of the connected component it belongs to. This representation compresses the image from as many cells as the grid has to a small set of point coordinates plus marks, and all four PointSSIM measures (intensity, area coverage, heterogeneity, and spatial irregularity) are computed from it.","core_discovery":"The central claim is that the structure of a binary image can be captured by the locally adaptive maxima of its minimal distance transform, and that comparing the resulting marked point processes is a valid, resolution-invariant way to compare images. The paper argues that this representation bypasses pixel-to-pixel correspondence, which is what makes pixel-based metrics like MSE and SSIM resolution-dependent. PointSSIM is defined as the normalized inverse Euclidean distance between the four summary vectors of the two images; a value near 1 means the images have nearly identical point-process summaries, near 0 means they differ strongly. The paper reports that on five synthetic classes of binary images, PointSSIM separates the classes more sharply than SSIM, MSE, or MS-SSIM, and that its measures remain largely stable when the same scene is generated at 256x256, 512x512, and 1024x1024 resolutions.","pith_inferences":["Because the anchor points are essentially skeletal points of the distance transform, PointSSIM is measuring a scale-space skeleton summary; it could be viewed as a low-dimensional signature of the image's medial axis, a connection the paper does not draw.","The same anchor-point machinery could be generalized to grayscale images by replacing the binary distance transform with a morphological gradient or intensity ridge detection; the paper only suggests gradients in passing, so this is an extension, not a claim.","A likely failure mode beyond the paper's own note: images that differ only by curvature or fine texture will be conflated, since the four measures ignore local shape; a test would be PointSSIM between straight and curved objects of equal area, which the paper's own figures suggest are not separated.","The variance-irregularity measure V4 uses 100 square subregions; the choice of grid partition is fixed, and a resolution change alters the physical size of these subregions, which may itself shift V4 even without smoothing effects."],"forward_implications":["Two binary images no longer need the same pixel dimensions: PointSSIM gives a similarity score for any pair of resolutions, e.g., 256x256 against 1024x1024.","Rotation of the input image leaves the four measures essentially unchanged, so comparisons do not require alignment.","The four-measure vector can be used as a cheap, interpretable descriptor for classification or clustering of binary images.","PointSSIM can be plugged into generative-model training as a structural regularization term alongside pixel-level losses.","The method extends naturally to any task where binary maps from different sources or grid spacings must be compared, such as geological realizations."],"supporting_citations":[{"why":"Supplies the structural similarity index that PointSSIM is named after and whose luminance-contrast-structure comparison PointSSIM replaces with point-process measures.","marker":"Wang et al. (2004)"},{"why":"Establishes mathematical morphology as the conceptual source for anchors and invariant transformations.","marker":"Matheron & Serra (2000)"},{"why":"Defines the minimal distance transform whose local maxima form the anchor points.","marker":"Banerji (2000)"},{"why":"Contributes the notion of anchors of morphological operators, which the adaptive anchor-point extraction adapts.","marker":"Van Droogenbroeck (2009)"},{"why":"Provides the marked point-process framework used to represent images as points with radius and label marks.","marker":"Ripley (2014)"},{"why":"Gives the Poisson point-process variance formula that measure V4 compares against to assess clustering.","marker":"Daley & Vere-Jones (1990)"},{"why":"Supplies the connected-component labeling routine used to assign object labels to anchor points.","marker":"Virtanen et al. (2020)"}],"fun_headline_variants":["Binary images compared via anchor points, scale-free","PointSSIM: four numbers beat pixels at any resolution","Distance transform maxima power resolution-invariant metric","No pixel matching: summary vectors compare binary scenes","Small summary vector locks image similarity across scales"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole resolution-invariance argument depends on the anchor points of an object being essentially unchanged when the image is rescaled; but low-resolution images merge nearby objects and produce fewer local maxima, so the count-based measures V1 and V3 drift.","fun_headline_variants_meta":{"raw":{"variants":["Binary images compared via anchor points, scale-free","PointSSIM: four numbers beat pixels at any resolution","Distance transform maxima power resolution-invariant metric","No pixel matching: summary vectors compare binary scenes","Small summary vector locks image similarity across scales"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1152,"prompt_tokens":816,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":265}},"tokens_in":432,"tokens_out":336,"duration_ms":4912,"temperature":1.0,"reasoning_tokens":265,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:30:09.095950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a binary image containing two distinct objects separated by a one-pixel gap and compare PointSSIM for the same scene rendered at 256x256 and at 1024x1024. If the gap closes at low resolution, the low-resolution version has one connected object instead of two, so V3 (anchor points per object) roughly doubles or halves, and the scatter point falls far off the identity line. A single such image pair where V1 or V3 deviates by a large margin between resolutions would falsify the resolution-invariance claim in a concrete, checkable case.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the minimal distance transform whose local maxima form the anchor points."}],"review_version":1}