{"id":"c1e5f211-3242-4697-aa9e-7d6d32e90aa8","arxiv_id":"2501.00836","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A style extrapolation module followed by a transfer-learned classifier achieves state-of-the-art artistic style recognition on archaeological image fragments, validated on a new Pompeii fresco fragment dataset.","lead":"This paper presents a two-stage deep learning pipeline that first extrapolates the artistic style of an archaeological fragment into its empty background, then classifies the fragment's style with a fine-tuned network. It reports state-of-the-art accuracy on an existing fresco fragment benchmark and introduces a new Pompeii fresco fragment dataset with four controlled fragmentation geometries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"POMPAAF train/test split is unspecified; if fragments from the same source fresco appear in both sets, the reported accuracy gains reflect source-image memorization rather than style generalization.","rationale":"The paper's central claim is that the proposed deep style extrapolation framework achieves SOTA on fragment style classification. The most direct evidence is the POMPAAF benchmark, which reports large gains (up to 0.967 accuracy) over simple transfer learning. However, this evidence depends entirely on an undisclosed train/test split. Because POMPAAF is generated by fragmenting a modest number of source fresco images, a random per-fragment split would place many fragments from the same source wall in both training and test sets. Since fragments of the same wall share distinctive colors, textures, and damage patterns, the network can achieve high test accuracy by recognizing the source image rather than the artistic style. This would undermine the POMPAAF results and the associated analysis of fragment geometry. I agree with the reader's weakest assumption; no other concern is as directly load-bearing. The CLEOPATRA result alone is too weak to carry the SOTA claim: with only 80 test fragments, the 0.475 versus 0.406 accuracy gap is within sampling error. The appropriate disposition is to require the authors to disclose the split and re-evaluate under a group split by source fresco. Thus the reader's CONDITIONAL verdict stands, pending that test.","tokens_in":11127,"tokens_out":5291,"duration_ms":52495,"concrete_test":"Download the released POMPAAF data and code, and inspect the exact split indices to check whether any source fresco contributes fragments to both train and test. Then re-run the proposed pipeline under a strict group split, assigning all fragments of each source fresco to a single fold (e.g., stratified 80/20 by source image) and recomputing Table 2-5 accuracies. If the gap between Proposed and TL shrinks substantially or the high accuracies (0.846-0.967) drop by more than a few points, the reported gains are largely due to source leakage rather than style recognition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 reports POMPAAF results in Tables 2-5 without ever stating how the train/test split was performed. The dataset (Section 3.2) is constructed by fragmenting 311 source fresco images; each source yields multiple fragments per fragmentation method and piece count. If the split is per-fragment rather than per-source-fresco, then training and test fragments from the same wall share the same color palette, texture, lighting, and condition. A classifier can then memorize source-image identity and map it to the style label, reporting high accuracy on the test set without learning any generalizable notion of Pompeian style. This is not a minor implementation detail: the POMPAAF tables are the primary evidence for the claimed SOTA on 'varying styles and geometries,' and the accuracy-versus-#pieces trend in Section 5.3 is confounded by this leakage. The issue is distinct from the method's internal validity and cannot be resolved from the paper as written, since no split protocol, number of training/test fragments, or class-balance statistics are provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for classifying the artistic style of image fragments: a \"style extrapolation\" preprocessing module, implemented as a ResNet18-based autoencoder with a residual connection and trained with a self-referential style loss (Eq. 3) plus a masked content loss (Eq. 4), followed by a fine-tuned EfficientNet classifier. The method is evaluated on the existing CLEOPATRA benchmark and on POMPAAF, a new dataset synthesized by fragmenting 311 Pompeian fresco images with four fragmentation methods at several piece counts. The authors report accuracy of 0.475 on the CLEOPATRA test set versus 0.406 for transfer learning alone and 0.28 for the prior random-forest model, and up to 0.967 accuracy on the square-fragment POMPAAF subset. The central claim is that the style extrapolation module improves fragment style classification over transfer learning and prior art.","tokens_in":11258,"tokens_out":5104,"duration_ms":51316,"significance":"If the reported results are reproducible, the work offers a useful preprocessing idea and a new public benchmark for a niche but practically relevant task. The POMPAAF dataset and code release are concrete contributions that could support future comparisons. The main claims, however, are entirely empirical and currently rest on evaluation details that are not fully documented. The paper contains no theoretical or machine-checkable results; its value hinges on whether the reported accuracy gains are genuine and not an artifact of data leakage or of inadequate baselines.","major_comments":[{"comment":"The paper never states how POMPAAF was partitioned into training and test sets. Since the dataset is generated by fragmenting 311 source fresco images, a per-fragment random split can place fragments from the same wall in both train and test. A classifier could then memorize per-fresco color, texture, and condition cues and inflate accuracy without learning a generalizable notion of style. This is load-bearing because all POMPAAF results are the main evidence for the claimed advantage on 'varying styles and geometries.' Please specify whether the split is by source fresco image or by individual fragment, report test-set sizes and class balance, and if the current split is per-fragment, re-run the experiments with a source-disjoint split.","section":"Section 5.3, Tables 2-5"},{"comment":"No experiment is repeated or reported with variance. The CLEOPATRA test set has only 80 pieces, so the 0.475 versus 0.406 gap may be within random noise, and the POMPAAF tables report single runs without error bars. Please provide confidence intervals, repeated-seed results, or statistical significance tests, and report per-class metrics for POMPAAF, especially for the small-piece regimes.","section":"Section 5.2 and 5.3"},{"comment":"The weights lambda and mu in the combined style extrapolation loss LSE = lambda*LAS + mu*LMC are hand-chosen, but no values are given, and no sensitivity analysis is reported. Together with the missing training set, epoch count, optimizer, and learning rate for the extrapolator, this makes the method not reproducible as written. Please add these details.","section":"Section 4.1, Eq. (5)"},{"comment":"Because Eq. (3) forces the output style to match the input style and the residual connection sums the input and the network output, an identity mapping is a trivial solution that satisfies both losses (the content loss is masked to fragment pixels). The paper provides no example outputs or quantitative evidence that the extrapolator actually fills background regions rather than degenerating. Please report qualitative visualizations of extrapolated fragments and a quantitative analysis of how much the background is modified.","section":"Section 4.1, Eqs. (3)-(5)"},{"comment":"The transfer-learning baseline is compared against the proposed method, but it is unclear whether the baseline receives the same input size, background handling, and preprocessing as the proposed method. Without an ablation that replaces the extrapolator with simple padding, cropping, or inpainting to the classifier's input size, the observed accuracy gains cannot be attributed to the extrapolation module rather than to a more favorable input representation.","section":"Section 4.3 and 5.3"}],"minor_comments":[{"comment":"There are numerous typos, including 'Forbenius' in Eq. (1), 'frgments' in Fig. 2, 'corssing-cuts' in Section 5.3, 'Pomepian' in Section 3, and 'CLEOPARTA' in Section 6. Please proofread carefully.","section":"Throughout"},{"comment":"The 12-piece Precision row is written as '0.45' while all other entries use three decimals; please format all table entries consistently.","section":"Table 4"},{"comment":"Eq. (7) defines accuracy through a binary confusion matrix, but the evaluation is multi-class. Please clarify whether the reported accuracy is overall accuracy or per-class accuracy and whether any macro-averaging is applied to accuracy itself.","section":"Section 5.1, Eq. (7)"},{"comment":"The comparison with Cascone et al. relies on numbers from the original paper. Please state explicitly whether the same CLEOPATRA train/test split and input preprocessing were used, and whether the random-forest baseline was re-run or only cited.","section":"Section 5.2, Table 1"},{"comment":"The footnote says recall was calculated from precision and F1. For macro-averaged metrics this inference is only valid if the same averaging was used; please confirm the calculation.","section":"Table 1 footnote"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible but currently under-supported by the missing split protocol and the absence of error bars. The POMPAAF leakage concern is the most serious issue: if the released benchmark has no source-disjoint splits, its value for the community is limited. The paper is probably not ready for acceptance at a top venue, but the requested additions are feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a worthwhile paper with a real idea and a useful new dataset, but the POMPAAF experiments are missing a key detail—how the train/test split was made—and until that's clear, the headline results on that dataset are not fully trustworthy.\n\nWhat's actually new: the style extrapolation module. Instead of padding or cropping a fragment, they train an autoencoder to 'diffuse' the fragment's style into the full image, using a Gatys-style Gram matrix loss against the input and a masked content loss to preserve the fragment pixels. That's a simple, sensible idea, and I don't recall seeing it as a preprocessing step for fragment classification. They also introduce POMPAAF, a four-style fresco fragment dataset with four controlled fragmentation geometries, and they ship code and data. The analysis of fragment size and geometry effects is a nice addition.\n\nThe experiments support a modest claim: their method beats a fine-tuned EfficientNet and a simple CNN on both CLEOPATRA and POMPAAF. The gap is consistent, and the CLEOPATRA result (0.475 vs 0.406) is credible because that dataset comes with its own test set.\n\nThe main soft spot is the unknown split protocol for POMPAAF. The paper never says whether fragments from the same source fresco are kept together in train or test. If the split is per-fragment, training and test fragments from the same wall share color palette, lighting, and wear, and a classifier can memorize source identity rather than learn style. That would inflate every POMPAAF number, and the accuracy-versus-#pieces trend would be confounded. This is load-bearing. The authors need to state the split and ideally provide a per-fresco split as a control.\n\nTwo smaller issues: there's no ablation against simple padding or cropping to fill the canvas, so we don't know how much of the gain comes from the learned extrapolation versus just removing the background shape. And there are no error bars or repeated runs, though that's common for this kind of paper.\n\nOverall: the idea is genuine, the dataset is a contribution, but the validation needs one major fix. I'd send it to review, with the split question as the first request.","headline":"Genuinely useful idea and dataset, but the missing POMPAAF split protocol is a load-bearing hole in the experimental claims.","tokens_in":11842,"tokens_out":3765,"would_cite":true,"duration_ms":32013,"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 style extrapolation preprocessing step lets fine-tuned deep networks classify artistic style of image fragments, beating prior methods on CLEOPATRA and a new Pompeii benchmark.","keywords":["image classification","artistic style recognition","fresco fragments","style transfer","deep learning","POMPAAF dataset","cultural heritage","EfficientNet"],"falsifier":"Inspect the POMPAAF split: if the train and test sets share fragments cut from the same source fresco, retrain the pipeline with a fresco-disjoint split and compare accuracies; a large drop would indicate the model memorized individual frescos instead of learning style, while holding out entire walls and testing whether accuracy stays above chance would settle the question directly.","tokens_in":10876,"feed_emoji":"🏛️","tokens_out":5103,"duration_ms":44325,"temperature":0.7,"pith_summary":"The paper argues that a two-stage deep learning pipeline can identify the artistic style of small broken image fragments, a task that stumps non-expert humans and prior machine learning methods. The first stage extrapolates the fragment's style outward to fill the whole image frame with a stylistically consistent synthetic picture, and the second stage classifies that filled image with a fine-tuned EfficientNet. On the CLEOPATRA benchmark, the method reaches 0.475 accuracy versus 0.406 for transfer learning alone and 0.28 for the prior random forest; on the new POMPAAF dataset, it reaches up to 0.967 accuracy on square fragments. If correct, the approach would make automated re-sorting of mixed fresco fragments from archaeological sites more reliable.","feed_headline":"Style extrapolation classifies broken fresco styles to 96.7% accuracy","feed_subtitle":"A two-stage deep pipeline beats transfer learning on mixed Pompeii fragments, aiding reconstruction of ancient artworks.","key_machinery":"The style extrapolation module is the load-bearing component. It is a modified autoencoder with a pre-trained ResNet18 encoder, a convolutional decoder, and a residual connection, optimized by the loss $L_{\\mathrm{SE}} = \\lambda L_{\\mathrm{AS}} + \\mu L_{\\mathrm{MC}}$. Here $L_{\\mathrm{AS}}$ is the auto-style loss comparing Gram matrices of VGG activations of the input and output, and $L_{\\mathrm{MC}}$ is a masked content loss that forces the output to stay close to the input on fragment pixels while allowing the background to be filled. This turns an irregular fragment with an arbitrary background into a stylistically consistent full-frame image, which the classifier can then process without shape confounding.","core_discovery":"The central claim is that a style extrapolation module—a convolutional autoencoder trained with a Gram-matrix style loss and a masked content loss—can diffuse a fragment's stylistic features into the non-pictorial background, so a subsequent classifier sees a complete image rather than an arbitrary silhouette. The paper reports state-of-the-art results with this module feeding a fine-tuned EfficientNet: 0.475 accuracy on CLEOPATRA's 11 styles and between 0.846 and 0.967 accuracy across the four POMPAAF fragmentation geometries, always beating both the simple CNN baseline and transfer learning alone. Smaller and more numerous fragments consistently yield higher accuracy than larger pieces, which the authors attribute to reduced downscaling at a fixed input resolution. The paper also introduces POMPAAF, a dataset of 311 Pompeian fresco images artificially broken into four shape families.","pith_inferences":["Editorial inference: the paper never describes the POMPAAF train/test split protocol; if fragments from the same fresco appear in both sets, the classifier could memorize per-fresco colors and textures, inflating the reported accuracies.","Editorial inference: a fresco-disjoint split—training on some walls and testing on fragments from unseen walls—would directly test whether the model learns a general Pompeian style rather than instance-specific appearance.","Editorial inference: coupling the style extrapolation idea with diffusion-based inpainting or attention-based classifiers, which the paper names as future directions, could push accuracy higher or reduce data requirements, though this is not tested here.","Editorial inference: because CLEOPATRA's classes are historical eras, the same pipeline might transfer to other fragment domains such as pottery or manuscript fragments, but the paper provides no evidence for that transfer."],"forward_implications":["On the CLEOPATRA dataset, the full pipeline's 0.475 accuracy beats transfer learning alone (0.406) and the prior random forest (0.28), indicating that the style extrapolation step is the source of the gain.","On POMPAAF, the method exceeds 0.94 accuracy for 160-piece square and non-convex fragment sets, suggesting fragment shape is not a barrier once style is extrapolated.","More fragments per fresco helps accuracy: for square fragments, accuracy rises from 0.846 at 12 pieces to 0.967 at 160 pieces, so smaller pieces do not lose information at a fixed input resolution.","Across crossing-cuts, non-convex, and eroded Voronoi fragment geometries, the proposed method outperforms both baselines by substantial margins, indicating the approach generalizes beyond square pieces.","The introduced POMPAAF dataset provides a new controlled benchmark for fragment style classification with varied fragment geometry."],"supporting_citations":[{"why":"Supplies the CLEOPATRA dataset and the prior state-of-the-art random forest result that the proposed method must beat.","marker":"[6]"},{"why":"Introduces neural style transfer and the Gram-matrix style loss on which the auto-style loss is built.","marker":"[12]"},{"why":"Provides the EfficientNet backbone used as the style classification module.","marker":"[32]"},{"why":"Provides the ResNet18 architecture used as the encoder of the style extrapolation module.","marker":"[14]"},{"why":"Supplies the Pompeii archive fresco images from which the POMPAAF dataset is created.","marker":"[24]"},{"why":"Defines the random forest model that is the prior baseline on the CLEOPATRA dataset.","marker":"[4]"},{"why":"Represents the earlier computational study of fresco fragment style classification that this work extends.","marker":"[5]"},{"why":"Provides the DAFNE fresco fragment dataset that inspired the eroded Voronoi fragmentation method.","marker":"[10]"}],"fun_headline_variants":["Pompeii fragments: style extrapolation boosts classification","Deep style extrapolation rebuilds fresco shards for 96.7%","AI completes broken fresco pieces to identify artistic style","Fragment to full image: style extrapolation improves accuracy","Cracked fresco styles recognized via deep extrapolation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported POMPAAF accuracies assume that no fragments from the same source fresco appear in both the training and test sets, since the paper does not describe the split protocol and the model could otherwise memorize each fresco's distinctive appearance rather than learn a general artistic style.","fun_headline_variants_meta":{"raw":{"variants":["Pompeii fragments: style extrapolation boosts classification","Deep style extrapolation rebuilds fresco shards for 96.7%","AI completes broken fresco pieces to identify artistic style","Fragment to full image: style extrapolation improves accuracy","Cracked fresco styles recognized via deep extrapolation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1489,"prompt_tokens":828,"completion_tokens":661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":580}},"tokens_in":444,"tokens_out":661,"duration_ms":7152,"temperature":1.0,"reasoning_tokens":580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:41:04.344331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the POMPAAF split: if the train and test sets share fragments cut from the same source fresco, retrain the pipeline with a fresco-disjoint split and compare accuracies; a large drop would indicate the model memorized individual frescos instead of learning style, while holding out entire walls and testing whether accuracy stays above chance would settle the question directly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ResNet18 architecture used as the encoder of the style extrapolation module."},{"cited_title":"In: Proceedings of Pictura Parietum","cited_arxiv_id":null,"evidence_quote":"Supplies the Pompeii archive fresco images from which the POMPAAF dataset is created."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the EfficientNet backbone used as the style classification module."},{"cited_title":"Journal of Ambient Intelligence and Humanized Com- puting 14(4), 4087–4097 (2023) 2, 3, 4, 7, 9, 10, 11, 12","cited_arxiv_id":null,"evidence_quote":"Supplies the CLEOPATRA dataset and the prior state-of-the-art random forest result that the proposed method must beat."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Introduces neural style transfer and the Gram-matrix style loss on which the auto-style loss is built."},{"cited_title":"Machine learning45, 5–32 (2001) 3, 10","cited_arxiv_id":null,"evidence_quote":"Defines the random forest model that is the prior baseline on the CLEOPATRA dataset."},{"cited_title":"In: International Conference on Image Analysis and Processing","cited_arxiv_id":null,"evidence_quote":"Represents the earlier computational study of fresco fragment style classification that this work extends."},{"cited_title":"Pattern Recognition Letters138, 631–637 (2020) 3, 6","cited_arxiv_id":null,"evidence_quote":"Provides the DAFNE fresco fragment dataset that inspired the eroded Voronoi fragmentation method."}],"review_version":1}