{"id":"dd3a9d8c-20e7-40c0-a923-c2e66fc9f54f","arxiv_id":"2502.01186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An SSIM-based distance, computed after SIFT alignment, separates same-die from different-die coin pairs nearly perfectly on a newly released 329-image dataset.","lead":"This paper introduces a scoring method that tells whether two ancient coins were struck from the same engraved die by aligning their photographs and comparing overall structure, and it releases the first public labeled dataset of 329 coin images for that task. The method is fast and reaches near-perfect accuracy on the test collections, which could replace many hours of manual die-link checking by archaeologists.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The comparative claim rests on a self-described 'inspired by' baseline that matches none of the cited pipelines exactly, so 'outperforms current techniques' is not yet demonstrated.","rationale":"The reader's weakest_assumption is exactly the load-bearing point: correctness of the comparative claim depends on Algorithm 2 faithfully representing the current state of the art. My reading of Sections 3 and 4 confirms that Algorithm 2 is a hybrid approximation rather than a reproduction of any of the three cited methods. This is not an internal inconsistency or a disagreement with consensus; it is a mismatch between the claim ('outperforming current techniques') and the evidence (a comparison against a custom baseline). The concern is concrete and testable: original code or exact re-implementations would settle it. The paper has real independent support: a publicly released labeled dataset, explicit preprocessing and algorithm descriptions, default parameters, and reproducible tables. The SSIM distance itself may well be useful, and the clustering results on the provided data are suggestive. However, the central comparative claim is not yet verified. Since the reader already issued CONDITIONAL, my analysis does not move the verdict; it reinforces that condition. I would not escalate to REJECT because the dataset and method contributions stand on their own and the baseline problem is fixable with additional experiments or softened claims. I would not downgrade to UNVERDICTED because the paper is empirically substantial and the specific missing comparison is well-defined. Thus the appropriate verdict remains CONDITIONAL, with the condition being that the comparison be validated against actual prior pipelines.","tokens_in":17100,"tokens_out":4405,"duration_ms":48026,"concrete_test":"Run each original pipeline from [46], [21], and [30] on the released 329-image benchmark, either by obtaining the authors' code or by implementing the exact descriptions in Section 3 (ORB+brute-force averaging for [46]; GP+VLFeat+bounded-distortion+Procrustes for [21]; SIFT+ratio test+bounded-distortion+combined distance for [30]). Then compare per-dataset ROC AUC, PR AUC, and ACmax ARI with the SSIM-based distance and with Algorithm 2. If any original method differs materially from Algorithm 2 (for example, by more than 0.05 ROC AUC on DS3 or DS8), the headline comparison is unsupported and must be re-run against the true methods. As a secondary check, ablate Algorithm 2 component-by-component (VLFeat versus ORB descriptors, bounded-distortion versus RANSAC matching, and the log(P)+1/nin combination) to identify which proxy choices drive the apparent gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is comparative: the SSIM-based distance outperforms current die-link techniques. The evidence compares against Algorithm 2, called 'Procrustes-based', which the text itself describes as 'the best possible pipeline inspired by [21,30,46]' and which is therefore 'referred to as Procrustes-based'. This is not a faithful reproduction of any cited method. According to Section 3, [46] uses ORB keypoints with averaged brute-force descriptor distances; [21] uses Gaussian-process keypoints with VLFeat descriptors, bounded-distortion feature matching, and a distance combining Procrustes error with match count; [30] uses SIFT with ratio test and bounded-distortion matching, combining Procrustes distance, match count, descriptors, and average local gradients. Algorithm 2 instead combines GP keypoints with ORB descriptors, uses brute-force cross-check matching, applies a RANSAC homography, and defines dP(A,B)=log(P)+1/nin, a formula matching none of the three cited distances. No ablation or released code shows that this proxy reproduces the behavior of the original methods. Therefore the empirical superiority of SSIM over 'current techniques' is not established; at most, SSIM beats a custom hybrid baseline. The claim could survive if the proxy is shown equivalent, but that premise is currently unsupported. The dataset and the SSIM distance itself remain useful contributions independent of this issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a labeled dataset of 329 ancient coin images for die link detection, proposes an SSIM-based image distance (Algorithm 1) that aligns images with SIFT and then computes a metric derived from the SSIM index, and evaluates this distance with agglomerative clustering using leave-one-out threshold estimation on eight datasets from the Juillac hoard. The authors claim that the SSIM-based distance achieves near-perfect die link identification when combined with a max-threshold aggregation rule (ACmax), and that it outperforms current techniques used in the field. The paper also reports comparisons against FSIM-, Procrustes-, and VGG-based distances.","tokens_in":17402,"tokens_out":5111,"duration_ms":50950,"significance":"If the central claims are substantiated, the released dataset and the fast, parameter-free SSIM distance would be valuable contributions to computational numismatics, providing a benchmark and a practical pre-screening tool. The evaluation is thorough in several respects: leave-one-out threshold estimation, multiple clustering and classification metrics, comparisons against FSIM and VGG baselines, and use of default library parameters for the proposed method. However, the comparative claim against the state of the art rests on a baseline that the authors themselves describe as 'inspired by' published methods rather than a faithful reproduction, so the empirical superiority over existing techniques is not yet established. The dataset itself is also filtered by lighting conditions without an analysis of the excluded images, which weakens the benchmark's representativeness.","major_comments":[{"comment":"The baseline used for the comparative claim is not a faithful implementation of any cited method. Section 3 describes [21] as using VLFeat descriptors with bounded-distortion matching, [30] as using SIFT with ratio test and bounded-distortion matching, and [46] as averaging ORB descriptor distances. Algorithm 2 instead combines Gaussian-process keypoints with ORB descriptors, cross-check matching, a RANSAC homography, and the distance dP(A,B)=log(P)+1/nin, which does not match any of the three cited pipelines. Since the paper's headline claim is that the SSIM-based method outperforms current techniques, the experiments need to compare against the actual published methods, or the authors must release the proxy code and demonstrate that it reproduces the original methods' behavior. In the absence of that, the claim should be restricted to the implemented baseline.","section":"Section 4, Algorithm 2, and Section 5"},{"comment":"The paper excludes 72 of 401 images because they are not lit from the left, but it provides no analysis of these excluded images or of how their removal affects the reported results. Because the dataset is presented as a benchmark and the exclusion criterion relies on the photographer's setup rather than an objective image property, the authors should report how many die-linked pairs were excluded and whether the remaining data are representative. At a minimum, the potential selection bias should be discussed as a limitation of the dataset.","section":"Section 2"},{"comment":"The claim of 'near-perfect die link identification' depends on the ACmax aggregation rule, whose choice is justified only post hoc from the histograms in Fig. 7. The leave-one-out threshold estimation is performed on the other datasets from the same hoard with the same lighting and preprocessing, so the threshold may not generalize to other coin collections. Please provide an independent validation or a more principled justification for selecting the maximum of the optimal thresholds, and discuss the variability of results across aggregation rules as part of the central claim.","section":"Section 5, Table 4"}],"minor_comments":[{"comment":"The caption states 'Procruste-base distance in blue' but the SSIM-based distance is also described as blue in the same caption; the colors should be disambiguated.","section":"Figure 6 caption"},{"comment":"There is a typo: 'inpired' should be 'inspired' in the sentence 'Now that the baseline distance inpired by the state of the art methods...'","section":"Section 4, paragraph after Algorithm 2"},{"comment":"The column header 'Ajusted Rand Index' should be 'Adjusted Rand Index'.","section":"Table 4"},{"comment":"The syntax 'if|s− 1|> 0.25; /* Wrong estimation */ then' contains an extra semicolon and an awkward comment placement; this should be rewritten for clarity.","section":"Algorithm 1, line 9"},{"comment":"The sentence 'Out of the eight datasets we kept, i.e. 401 coins, the lighting was correct for only 329' is confusing because it suggests the datasets originally contained 401 coins and 329 were kept; it would be clearer to state that 72 images were removed due to lighting.","section":"Section 2"},{"comment":"The statement that k-means, k-medoids, and CoRe without repulsion 'resulted in very poor clustering predictions' is not accompanied by results; including these numbers in the supplementary material would strengthen the evaluation.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for a computer vision venue with applied cultural heritage focus. The main technical concern is the baseline fidelity, which is acknowledged by the authors but not resolved; the paper's comparative claim cannot be accepted as stated. The dataset release is a positive feature, and the SSIM-based method itself appears sound and fast. The exclusion of 72 images and the post hoc selection of ACmax also merit attention but are fixable with additional analysis or softened claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the paper's durable value is the dataset. It releases 329 labeled coin images with ground-truth die links, the first public resource for this task, and that alone makes it worth a look. The SSIM-after-SIFT distance is simple, fast, and on its own data it separates linked from unlinked pairs cleanly, with no fitted constants and default library parameters. Those are real strengths. The comparative claim, however, is soft: the 'Procrustes-based' baseline is explicitly described as 'the best possible pipeline inspired by [21,30,46]', and its components (GP keypoints, ORB descriptors, brute-force cross-check, RANSAC homography, log(P)+1/nin) match none of the three cited methods exactly. So the headline that SSIM 'outperforms current techniques' is really only evidence that it beats a custom hybrid. That does not sink the paper, but it does mean the central comparative claim is overstated as written. The authors should either run the original implementations where code is available, or at least show that their proxy reproduces the original methods' behavior on some validation set. The 72 excluded images due to lighting are also a genuine caveat; no evaluation on them is reported, and the exclusion could bias results if lighting correlates with die condition. The near-perfect ARI numbers come from very small positive-link counts (DS6 has a single link), so those scores are more illustrative than solid evidence. The choice of ACmax over the other threshold aggregation rules is reasonable but appears post-hoc; the authors give a plausible distribution-based justification, yet it is still a selection made after seeing the results. What holds up: the method itself is well-motivated, the metric choice (Brunet's SSIM distance) respects the triangle inequality, and the dataset is a real contribution that will help this small field move forward. I'd send it to review: the application is narrow but important to numismatics, and the dataset deserves public scrutiny. Ask for code release, a toned-down comparative claim, and at least a robustness check on the excluded images. I would not cite it in my own work, but someone working on coin imagery or small-scale image clustering should.","headline":"A genuinely useful first dataset and a credible screening distance, but the 'outperforms current techniques' claim is not yet demonstrated because the baseline is only inspired by prior work.","tokens_in":17915,"tokens_out":2166,"would_cite":false,"duration_ms":25004,"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":"A SSIM-based image distance can identify ancient coins struck from the same die, and clustering on that distance recovers the true die-link groups nearly perfectly.","keywords":["coin die link identification","SSIM","structural similarity","ancient numismatics","image distance","clustering","benchmark dataset","Juillac treasure"],"falsifier":"Re-run the same eight-dataset experiment with the original implementations of the cited methods (for example, the actual code from the papers cited as [21] and [30]) and with a second coin hoard photographed under varied lighting; if the SSIM-based distance then fails to separate true die links from non-links at the reported accuracy, the near-perfect clustering claim would not generalize.","tokens_in":16916,"feed_emoji":"🪙","tokens_out":2539,"duration_ms":29218,"temperature":0.7,"pith_summary":"This paper sets out to replace the slow, manual comparison of ancient coin photos with a fast automatic distance measure that tells when two coins were struck by the same engraved die. It introduces the first publicly labeled image dataset for die-link detection, containing 329 coin photos from the Juillac treasure, and proposes an SSIM-based distance computed after aligning coin images with keypoint matching. On this dataset, the distance separates true die links from non-links better than a Procrustes-based baseline inspired by prior work, and a simple agglomerative clustering with a max-threshold rule achieves perfect or near-perfect clustering (ARI 1.0 on six of eight datasets). If the result holds, numismatists can pre-screen huge hoards automatically and focus manual verification on the most likely links.","feed_headline":"SSIM score finds coin die links near-perfectly","feed_subtitle":"A fast full-image similarity distance plus smart thresholding could cut months of manual coin comparison work.","key_machinery":"The central object is the SSIM-based distance $M(A,B)$, a per-pixel average of the local metric $M_{ij}(A,B)=\\sqrt{2-l_{ij}^{AB}-s_{ij}^{AB}c_{ij}^{AB}}$, which combines luminance, contrast, and structure terms from the structural similarity index. It is a true metric (satisfies the triangle inequality), unlike a simple $1-\\text{SSIM}$ score, which matters for distance-based clustering. The pipeline uses preprocessing (grayscale, circular crop, CLAHE, non-local means denoising), then SIFT keypoints solely to estimate an affine overlay of the two images, after which the global SSIM distance is computed over the aligned surfaces. The baseline is a Procrustes-based distance that combines the log of the Procrustes distance between keypoint configurations with the inverse number of RANSAC inliers.","core_discovery":"The paper claims that a global structural-similarity distance, computed after affine alignment, is a better discriminator of coin die links than the keypoint-and-Procrustes approaches previously used in the field. The distance is defined from a local SSIM metric, $M_{ij}(A,B)=\\sqrt{2-l_{ij}^{AB}-s_{ij}^{AB}c_{ij}^{AB}}$, averaged over all pixels; the alignment is obtained by SIFT keypoints with a ratio test and an affine transform, with the transform discarded if the estimated scale is far from 1. Combined with agglomerative clustering using the maximum of leave-one-out optimal thresholds for the cutoff, the approach yields adjusted Rand index 1.0 on DS1, DS2, DS4, DS5, DS6, and DS7, and 0.947 and 0.909 on DS3 and DS8, with perfect precision on all datasets. The authors also release the dataset and position the method as a fast pre-screening tool for large coin hoards.","pith_inferences":["The core idea -- align globally, then compare with a full-image structural metric -- could be tested on other fine-grained object-identity problems, such as tool-mark matching in forensics or seal impressions in archaeology, where imaging conditions are similarly controlled.","The paper leaves implicit that its threshold-learning scheme assumes consistent lighting orientation (always lit from the left), so a natural extension is to test how the distance degrades when lighting direction varies or when coins are photographed under different protocols.","A straightforward improvement, not explored here, would be to replace the affine alignment with a full homography when scale is trusted, or to use multi-scale SSIM variants (MS-SSIM, CW-SSIM) that the authors list as future work; these could improve the two harder datasets DS3 and DS8.","The claimed superiority over state of the art rests on a re-implementation of prior methods rather than their original code; a stronger comparison would run the original published pipelines on the same dataset, which the authors note is currently impossible because those codes were not released."],"forward_implications":["If the SSIM distance performs as reported, coin die-link analysis of large hoards becomes practical: the 1,350-hour manual comparison in the L'Isle-Jourdain example could be reduced to an automatic pre-screen that ranks pairs for human verification.","The released dataset gives other researchers a common benchmark for die-link detection, enabling direct comparison of future methods rather than evaluation on private collections.","The success of the max-threshold aggregation strategy suggests that the inter-cluster distance distribution has a stable lower bound across datasets, so thresholds learned on one set transfer to another of the same imaging protocol.","Because the SSIM-based distance runs in hours versus days for the Procrustes baseline, the method is plausible for online production use on newly excavated or photographed coin collections.","The method's global nature means it exploits full image detail rather than only matched keypoints, which may generalize to other artifact-matching tasks where local features fail to capture subtle die-specific engravings."],"supporting_citations":[{"why":"Defines the structural similarity index that the new distance is built from.","marker":"[49]"},{"why":"Provides the local SSIM-based metric $M_{ij}$ that satisfies the triangle inequality, making the average a true distance.","marker":"[11]"},{"why":"SIFT keypoints and descriptors are used in the alignment step of the SSIM-based pipeline.","marker":"[26]"},{"why":"Represents the state-of-the-art die-link pipeline that the Procrustes baseline is inspired by.","marker":"[21]"},{"why":"Provides the CoRe Bayesian distance clustering baseline and a comparison method for the clustering evaluation.","marker":"[30]"},{"why":"The released dataset of 329 labeled coin images used for all experiments.","marker":"[24]"},{"why":"VGG11 pretrained features are used as a deep-learning baseline distance for comparison.","marker":"[44]"},{"why":"FSIM is used as an alternative full-reference similarity baseline that performs comparably but slower.","marker":"[51]"}],"fun_headline_variants":["SSIM score nails ancient coin die links near-perfectly","Near-perfect die link ID with new SSIM scoring method","SSIM-based scoring outperforms old coin die link techniques","First public coin dataset plus SSIM score hits perfect links"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison baseline is an approximation of prior methods re-implemented by the authors, not the original published code, so the claim that the SSIM distance outperforms current techniques depends on this proxy being faithful.","fun_headline_variants_meta":{"raw":{"variants":["SSIM score nails ancient coin die links near-perfectly","Near-perfect die link ID with new SSIM scoring method","SSIM-based scoring outperforms old coin die link techniques","First public coin dataset plus SSIM score hits perfect links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1323,"prompt_tokens":899,"completion_tokens":424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":515,"tokens_out":424,"duration_ms":4778,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:16:18.629922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same eight-dataset experiment with the original implementations of the cited methods (for example, the actual code from the papers cited as [21] and [30]) and with a second coin hoard photographed under varied lighting; if the SSIM-based distance then fails to separate true die links from non-links at the reported accuracy, the near-perfect clustering claim would not generalize.","supporting_citations":[{"cited_title":"Image quality assessment: from error visibility to structural similarity","cited_arxiv_id":null,"evidence_quote":"Defines the structural similarity index that the new distance is built from."},{"cited_title":"On the mathematical properties of the structural similarity index","cited_arxiv_id":null,"evidence_quote":"Provides the local SSIM-based metric $M_{ij}$ that satisfies the triangle inequality, making the average a true distance."},{"cited_title":"Object recognition from local scale-invariant features","cited_arxiv_id":null,"evidence_quote":"SIFT keypoints and descriptors are used in the alignment step of the SSIM-based pipeline."},{"cited_title":"Unsupervised Statistical Learning for Die Analysis in Ancient Numismatics","cited_arxiv_id":"2112.00290","evidence_quote":"Represents the state-of-the-art die-link pipeline that the Procrustes baseline is inspired by."},{"cited_title":"Cohesion and repulsion in bayesian distance clustering","cited_arxiv_id":null,"evidence_quote":"Provides the CoRe Bayesian distance clustering baseline and a comparison method for the clustering evaluation."},{"cited_title":"Datasets for Accadil (V1) , 2024","cited_arxiv_id":null,"evidence_quote":"The released dataset of 329 labeled coin images used for all experiments."},{"cited_title":"Fsim: A feature similarity index for image quality assessment","cited_arxiv_id":null,"evidence_quote":"FSIM is used as an alternative full-reference similarity baseline that performs comparably but slower."}],"review_version":1}