{"id":"a395918b-e5a4-4805-a4b5-917ae147c8cd","arxiv_id":"2508.18641","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A clustering-based contrastive loss that uses a clean oracle bone font library as an anchor improves character detection in rubbing images by small margins over three detection baselines.","lead":"Researchers added a clustering-based training loss to three object detection networks so that oracle bone characters are pulled toward clean font-library examples and pushed away from noise and cracks in feature space. The method reports modest gains in detection accuracy on two oracle bone inscription datasets, though some configurations show no gain or a small loss.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may reflect test-set hyperparameter tuning; no train/validation/test split is described for either dataset.","rationale":"The reader's weakest_assumption focuses on the comparability of OBC font-library features and real rubbing features, which is a valid mechanistic concern and is supported by negative results in Tables 5 and 9. However, I believe the single most load-bearing concern for the central claim is the absence of a described train/validation/test split combined with hyperparameter tuning (Bayesian optimization) that appears to be performed on the evaluation set. This directly threatens the validity of every reported improvement, because the numbers could be an artifact of selection. The reader's rationale does mention this issue ('hyperparameters appear to be selected on the evaluation data without a described split') and uses it to justify the CONDITIONAL verdict, so my conclusion is aligned with the verdict. I do not advocate changing the verdict because the concern is about missing evidence rather than a demonstrated failure. The proposed concrete test — re-running with a proper split and validation-based tuning — would settle whether the concern lands. If the gains persist under a clean protocol, the central claim would be substantially strengthened; if not, the verdict should move toward REJECT. For now, CONDITIONAL remains appropriate.","tokens_in":16601,"tokens_out":7008,"duration_ms":67325,"concrete_test":"Re-run the main comparisons on a properly defined split of each dataset (e.g., 80/10/10), tune all hyperparameters (OBC count, τ, M, N, λ1, λ2, λ3) exclusively on the validation split, and evaluate on the held-out test split with multiple training seeds for both baselines and the proposed method. If the average AP/AP50/F1 improvements over baselines fall below the reported margins or become negative, the central claim of significant, generalizable improvement is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on the numbers in Tables 2 and 3, but the paper never specifies how the OBIs detection dataset (9,500 images) or OBIMD (10,077 images) are split into training, validation, and test sets. The method introduces several hyperparameters — the number of OBC images (Table 5), temperature τ (Table 6), cluster counts M and N, and loss weights (Table 8) — and Table 7 reports results after 'Bayesian Optimization technique to tune our hyperparameters.' If this tuning was performed by monitoring performance on the same evaluation set used to report final AP/AP50/F1, the observed gains of +0.7 to +2.9 AP50 could be optimistically biased. Moreover, baseline results in Tables 2 and 3 appear to be single runs without variance, while Table 10 reports mean±std only for the proposed method and only across K-means initializations, not across training seeds or data splits. Without a clear evaluation protocol, the loading-bearing condition that the improvement is real and generalizable is not secured. This is not an accusation of misconduct; it is a missing methodological detail that prevents the reader from ruling out selection bias. The feature-comparability assumption identified by the reader is real, but it is secondary: even if the font anchor is imperfect, the method could still work; the empirical claim stands or falls on the integrity of the evaluation protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a clustering-based contrastive learning method for detecting Oracle Bone Inscriptions (OBIs) in rubbing images. The method extracts RoI features from three detection frameworks (Faster R-CNN, DETR, Sparse R-CNN), clusters negative features from background regions and positive features from the OBC font library, and adds a contrastive loss L_clus that pulls positive OBI RoI features toward the mean font cluster center and away from negative cluster centers. The total loss is L = λ1 L_clus + λ2 L_class + λ3 L_box. Experiments on two OBI datasets report improvements in AP, AP50, and F1-score over several baselines, with additional ablations on hyperparameters. The paper includes data and code availability links.","tokens_in":16790,"tokens_out":5722,"duration_ms":46827,"significance":"The central idea—using a clean, expert-curated font library as an anchor in feature space—is a reasonable and potentially useful approach for a domain with scarce clean data. If the claimed improvements are real and generalizable, the method would be a simple plug-in module for OBI detection and possibly other ancient-character recognition tasks. Strengths include end-to-end training without auxiliary networks, use of two real datasets, and public code/data links. However, the empirical case is currently weakened by the missing evaluation protocol, the absence of variance measures for baselines, and a contradiction between the abstract and Table 2.","major_comments":[{"comment":"The abstract states that 'all frameworks demonstrate significant performance improvements,' but Table 2 shows that Ours(Faster R-CNN-Swin) decreases AP by 0.2 (47.8 vs 48.0), AP50 by 0.2 (86.6 vs 86.8), and AP75 by 0.6 (48.2 vs 48.8) relative to baseline. The text in §3.2 concedes 'only the accuracy rate on Faster R-CNN-Swin has slightly decreased.' This contradiction must be resolved, and the word 'significant' should be removed or justified with statistical tests.","section":"Abstract, §3.2, Table 2"},{"comment":"The denominator in Eq. (2) sums exp(p_n · C_N / τ) over negative cluster centers C_N, but the surrounding prose states 'the denominator is the similarity of the feature point and all of its positive sample cluster centers.' This mismatch makes the loss definition ambiguous: if the denominator only contains negatives, the loss is not the standard InfoNCE form and its gradient behavior is unclear. Please correct the equation or the description, and clarify the summation index (which should not reuse n).","section":"§2.4, Eq. (2)"},{"comment":"The paper never specifies how the OBIs detection dataset (9,500 images) and OBIMD (10,077 images) are split into training, validation, and test sets. Table 7 reports results obtained after 'Bayesian Optimization technique to tune our hyperparameters,' but it is not stated whether the optimization monitored performance on the test set. If so, the reported gains could be optimistically biased by selection. The authors must describe the split, the tuning protocol, and whether the reported numbers were obtained on a held-out test set never used for model selection.","section":"§3.2, §3.3, Tables 2, 3, 7"},{"comment":"Baselines in Tables 2 and 3 are single runs with no error bars or significance tests, while Table 10 reports mean ± std only for the proposed method and only across K-means initializations. Since many reported gains are small (e.g., +0.7 AP for Sparse R-CNN), the reader cannot judge whether the improvements are meaningful. Please report variance or confidence intervals for the baselines as well, or at least provide a paired significance test over multiple training seeds and data splits.","section":"Tables 2, 3, 10"},{"comment":"The text says 'we conducted three experiments respectively on three models with the same parameters,' but Table 10 reports that experiments were 'repeated 5 times with different random initializations.' These statements are inconsistent. Also, varying only the K-means initialization does not test reproducibility across training seeds, data splits, or other stochastic elements; the claim 'The results are reproducible when rerunning the code' is therefore overstated.","section":"§3.3, Table 10"}],"minor_comments":[{"comment":"Typo: 'alse' should be 'also' in the sentence about OBIMD.","section":"§2.1"},{"comment":"Typo: 'F1-socre' should be 'F1-score' in the introductory sentence of the main results.","section":"§3.2"},{"comment":"The column header 'AR 50' should read 'AR50' for consistency with the text and other tables.","section":"Table 2"},{"comment":"Use the standard spelling 'RoIAlign' instead of 'RoI Align'.","section":"§2.3"},{"comment":"The notation {C_N} and {C_M} is ambiguous: it is not clear whether C_N denotes a single cluster center or the set of centers. Please distinguish between a cluster index and the set.","section":"§2.4"},{"comment":"The model named 'DENO-Swin' in Table 2 refers to DINO (per the text); please use consistent naming throughout.","section":"§3.2, Table 2"},{"comment":"The merged cells and overlapping values in Tables 8 and 9 are hard to read; please reformat them into separate rows with clear column boundaries.","section":"Tables 8, 9"},{"comment":"The sentence 'We'll conduct a contrastive learning of point and centers of clustering' is ungrammatical and should be rewritten.","section":"§2.4"},{"comment":"The evaluation metrics section does not define how F1-Score50 is computed from per-image predictions; please specify whether it is aggregated over all predictions or averaged per image with a confidence threshold.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. The central idea is modest but not unreasonable, and the font-library prior is externally grounded. The main risk is the evaluation protocol: without a documented train/validation/test split and with hyperparameters tuned via Bayesian optimization, the reader cannot rule out selection bias. The contradiction between the abstract and Table 2, and the mismatch in Eq. (2), must be fixed. I do not see a circularity problem with the method itself. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and worth taking seriously: use a clean, expert-curated oracle bone font library as a positive anchor and cluster centers of negative features as negatives in a contrastive loss on RoI features. That is a new combination for OBI detection, and it is refreshingly easy to bolt onto standard detectors. The gains in Tables 2 and 3 are small but mostly positive, and the paper does not oversell them in prose. The t-SNE visualizations actually show the intended separation. Good.\n\nThe soft spots are mostly about evidence quality. The paper never states how either dataset is split into train/val/test. Hyperparameters — OBC count, temperature, clusters, loss weights — are tuned via Bayesian optimization, and the reported test numbers look like they were selected on the same evaluation set. That makes the +1–3 AP50 claims vulnerable to selection bias. The paper even admits one negative result against a Swin backbone, which is fine, but it undercuts the abstract's “all frameworks demonstrate significant improvements.” The loss equation (Eq. 2) is also inconsistent with the prose: the denominator sums over negative centers only, not all positive centers, and the notation is sloppy. Table 10 reports variance only across K-means initializations, not training seeds or data splits, so reproducibility across runs is not demonstrated. The code and data links are there, but no commit hash or environment — so only principal reproducibility.\n\nOn the circularity concern: I agree with your low burden. The font library is external prior knowledge, and cluster centers come from training features, not from test labels. That part is clean. The feature-comparability assumption (do clean font glyphs live in the same feature space as degraded rubbing characters after the shared backbone?) is plausible, and the OBC=50 results suggest it can break, but the method still works when the count is sensible.\n\nWho is this for? People working on OBI detection or similar degraded-script tasks with a font library at hand. It is a modest, honest improvement, not a field-shaping result. The central claim is probably directionally right, but the missing evaluation protocol prevents a clean verdict.\n\nThe paper deserves a serious referee. I would send it to review, but require the authors to specify the train/val/test split, report variance across seeds, fix the equation/prose mismatch, and recalibrate the abstract. That is a solid conditional accept after revision.","headline":"A genuinely useful domain trick — font-library-anchored contrastive loss for OBI detection — but the empirical claim is undercut by missing split details and a weak equation.","tokens_in":17421,"tokens_out":607,"would_cite":true,"duration_ms":7301,"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":"Adding a clustering loss that anchors rubbing-image character features to clean font-library glyphs improves oracle bone inscription detection across Faster R-CNN, DETR, and Sparse R-CNN.","keywords":["oracle bone inscriptions detection","clustering-based representation learning","contrastive learning","font library prior knowledge","object detection","degraded archaeological imagery","K-Means clustering"],"falsifier":"Train the same three detectors with the OBC font-library images replaced by an equal number of clean glyph images from an unrelated script, keeping the clustering loss and all hyperparameters fixed; if average precision still rises, the oracle-font prior is not the active ingredient.","tokens_in":16301,"feed_emoji":"🏺","tokens_out":8561,"duration_ms":71795,"temperature":0.7,"pith_summary":"The paper tries to establish that a detection network can be made to separate genuine oracle bone characters from cracks, noise, and other look-alike structures by adding a clustering-based contrastive loss that pulls each character-region feature toward a clean, expert-drawn font-library glyph representation and pushes it away from negative cluster centers. The method needs no extra pre-trained network: font-library images and rubbing images pass through the same feature extractor during training, and the extra loss is simply added to the standard classification and box losses. If true, this means curated font libraries can act as reusable prior knowledge for detection in degraded archaeological imagery, improving the accuracy of three mainstream detectors (Faster R-CNN, DETR, Sparse R-CNN) on two oracle bone datasets with modest extra training cost. The paper's evidence is the consistent AP, AP50, and F1-score gains across frameworks and the feature-space visualizations showing character features become linearly separable from non-character features.","feed_headline":"Font glyph anchors improve oracle bone detection","feed_subtitle":"Pulling rubbing-image character features toward clean expert glyphs raises AP across Faster R-CNN, DETR, and Sparse R-CNN.","key_machinery":"The load-bearing mechanism is the clustering-based contrastive loss $\\mathcal{L}_{clus}$ computed on RoI feature vectors. During training, K-Means clusters the negative-anchor features from rubbing images into centers $\\{C_N\\}$ and clusters positive-anchor features from the OBC font-library images into centers $\\{C_M\\}$, then uses the average font center $C_M^{\\text{mean}}$ as the positive anchor for every sample feature. The loss is a softmax cross-entropy over similarities between the sample feature and the positive center versus all negative centers, with temperature $\\tau$ controlling how strongly hard negatives are penalized. The settings of $\\tau$, of the loss weight $\\lambda_1$, and of the number of font glyphs are model-specific, and the paper shows that too many font anchors (OBC=50) or a different clustering method (DBSCAN) can make the auxiliary loss harmful.","core_discovery":"The central discovery is that RoI features of oracle bone characters in rubbing images, which initially overlap with non-character structures, can be reorganized by an anchor-based contrastive loss that uses the OBC font library as a clean reference. The method clusters negative-anchor features into $N$ centers $C_N$ and positive font-glyph features into $M$ centers whose average is $C_M^{\\text{mean}}$, then applies the loss $\\mathcal{L}_{clus}(p_n) = -\\log \\frac{\\exp(p_n \\cdot C_M^{\\text{mean}}/\\tau)}{\\sum_n \\exp(p_n \\cdot C_N/\\tau)}$, pulling character features toward the average font-glyph center and away from non-character centers. This loss is added to the standard losses as $\\lambda_1 \\mathcal{L}_{clus} + \\lambda_2 \\mathcal{L}_{class} + \\lambda_3 \\mathcal{L}_{box}$, with the font images contributing only to $\\mathcal{L}_{clus}$. The paper reports that this reorganization improves average precision, AP50, and F1-score over the unmodified detectors on both the OBIs detection dataset and OBIMD, and its t-SNE visualizations show that character and non-character features become linearly separable.","pith_inferences":["The same anchor-cluster recipe could be tried for other degraded writing systems that have clean reference fonts, such as bronze inscriptions or seal script; the authors gesture at this generality but do not test it.","Because the loss collapses all positive font centers into one average point, it discards stylistic variation; a multi-positive or per-cluster contrastive form might preserve more information and could outperform the single-anchor version.","The mechanism predicts that gains will shrink as the input images get cleaner, since the separation the loss enforces has less room to help; this could be checked by adding synthetic clean vs. heavily degraded test splits.","Since the font images contribute only to the auxiliary loss, the idea could be grafted onto character recognition or retrieval tasks, turning a font library into a general feature regularizer."],"forward_implications":["Any proposal-based detector that produces RoI or query features can take the extra loss without architectural change, so the improvement should transfer to other instance-level detectors beyond the three tested.","Because the font-library prior compacts the character feature region and separates it from crack and noise structures, false positives under heavy degradation should drop.","For sparse-query detectors, smaller temperature values compensate for the smaller pool of negative samples and recover most of the gain.","The extra training cost is bounded and inference cost is unchanged, making the auxiliary loss a cheap add-on for production pipelines.","Using more font glyphs is not monotonically better; the best number of anchors is a tuned hyperparameter, and too many can degrade the model."],"supporting_citations":[{"why":"Supplies the clean OBC font-library glyph images whose features serve as the positive anchor center in the auxiliary loss.","marker":"[4]"},{"why":"The OBIs detection dataset used as the primary training and evaluation benchmark for baseline and proposed detectors.","marker":"[37]"},{"why":"The OBIMD multi-view dataset used as the second, independent evaluation benchmark.","marker":"[41]"},{"why":"Faster R-CNN is the dense-proposal baseline that the clustering loss is added to.","marker":"[38]"},{"why":"DETR is the transformer-based baseline used to test whether the loss transfers to query-based detectors.","marker":"[39]"},{"why":"Sparse R-CNN is the sparse-query baseline used to test the method's generality.","marker":"[40]"},{"why":"Provides the within-class clustering idea that the proposed contrastive loss extends to object detection.","marker":"[35]"}],"fun_headline_variants":["Clustering pulls oracle bone features toward font glyph centers","Contrastive loss with font library boosts oracle bone detection","Font-guided clustering sharpens oracle bone detection features","Oracle bone detection improved by clustering on font glyphs","Anchoring features to font glyphs lifts oracle bone detection AP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The clean, expert-written font-library glyphs must live close enough to the degraded rubbing characters in feature space that pulling the two together separates characters from noise instead of corrupting the detector.","fun_headline_variants_meta":{"raw":{"variants":["Clustering pulls oracle bone features toward font glyph centers","Contrastive loss with font library boosts oracle bone detection","Font-guided clustering sharpens oracle bone detection features","Oracle bone detection improved by clustering on font glyphs","Anchoring features to font glyphs lifts oracle bone detection AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1268,"prompt_tokens":950,"completion_tokens":318,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":239}},"tokens_in":566,"tokens_out":318,"duration_ms":3237,"temperature":1.0,"reasoning_tokens":239,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:54:14.602062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same three detectors with the OBC font-library images replaced by an equal number of clean glyph images from an unrelated script, keeping the clustering loss and all hyperparameters fixed; if average precision still rises, the oracle-font prior is not the active ingredient.","supporting_citations":[{"cited_title":"In: Journal of Physics: Conference Series, vol","cited_arxiv_id":null,"evidence_quote":"Supplies the clean OBC font-library glyph images whose features serve as the positive anchor center in the auxiliary loss."},{"cited_title":"In: Proceedings of the International Conference on Artificial Intelligence, Information Processing and Cloud Computing, pp","cited_arxiv_id":null,"evidence_quote":"The OBIs detection dataset used as the primary training and evaluation benchmark for baseline and proposed detectors."},{"cited_title":": Sparse r-cnn: End-to-end object detection with learnable proposals","cited_arxiv_id":null,"evidence_quote":"Sparse R-CNN is the sparse-query baseline used to test the method's generality."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Provides the within-class clustering idea that the proposed contrastive loss extends to object detection."}],"review_version":2}