{"id":"b4effed7-e499-4a64-98c9-378d49066713","arxiv_id":"2412.11668","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A temporal-frequency attention model and a 674,017-sample benchmark improve online Chinese writer retrieval.","lead":"This paper presents DOLPHIN, a neural network for retrieving handwriting samples written by the same person from short Chinese phrases, along with a new benchmark dataset called OLIWER. The authors report large gains over existing methods, and show that matching sampling rate and pressure information across datasets improves cross-domain retrieval.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparison may be undermined by the stated 2D-to-1D baseline adaptation: 'simply substituting the Conv2d module with the Conv1d module' (Sec. V-E) cannot faithfully preserve 2D-specific attention/pooling in CAL, OSNet, and CDNet, so DOLPHIN's large margins could be inflated.","rationale":"Reading the paper in good faith, the contributions are a new model (DOLPHIN) with temporal-frequency attention, and a large-scale dataset (OLIWER). The strongest claim is quantitative: DOLPHIN outperforms the best prior method by large margins (96.40 vs 93.32 R1; 69.83 vs 61.01 mAP on OLIWER). For this claim to hold, the baseline numbers must be trustworthy. The manuscript's description of baseline adaptation is the most fragile point: 'simply substituting the Conv2d module with the Conv1d module' cannot preserve the 2D-specific designs of CAL, OSNet, and CDNet. These methods are not generic classifiers; their performance depends on spatial attention, multi-scale 2D aggregation, or depth-space reasoning. A 1D analogue that keeps the same loss and training recipe is non-trivial and must be validated. If the released code reveals a literal substitution, the baselines are almost certainly broken; if it reveals a more thorough conversion, the description is misleading and the fairness of the resulting implementations remains unverified. Either way, the paper does not currently establish that the baselines are strong, so the reported margins may be inflated. This concern is more load-bearing than the dataset-traceability issue identified by the reader: while the undocumented 255 writers in DCOH-E are a reproducibility gap, it does not by itself invalidate the relative comparison on the same data; an unfair baseline does. The 30-vs-50 repetitions inconsistency compounds the reproducibility problem but is secondary. I recommend keeping the CONDITIONAL verdict: the authors should release faithful baseline implementations and rerun comparisons, or provide evidence that their adaptations preserve each method's core functionality. Until then, the empirical superiority claim is not fully supported.","tokens_in":23365,"tokens_out":6383,"duration_ms":60124,"concrete_test":"Obtain the released code (GitHub: SCUT-DLVCLab/DOLPHIN) and inspect the baseline conversion. Then re-run three strongest baselines—CAL, OSNet, CDNet—on OLIWER with a faithful 1D conversion that preserves each model's attention and pooling logic (e.g., keep 2D semantic operations over a height-1 spatial map, or implement the intended 1D analogue), using the same training protocol and losses. Compare Rank-1 and mAP to Table III. If any baseline's mAP increases by more than 2–3 points, DOLPHIN's margin is not robust; the comparison should be re-reported with properly adapted baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result—DOLPHIN surpassing CAL by 8.82 mAP on OLIWER and similar margins elsewhere—rests on a comparison against baselines adapted from 2D image models. Section V-E states: 'we convert them into their 1D version by simply substituting the Conv2d module with the Conv1d module in Pytorch implementations, which avoids altering their core designs and functionalities.' This is not a well-defined operation for the strongest baselines. CAL (ICCV'21) uses counterfactual attention computed over 2D spatial feature maps; OSNet (TPAMI'22) relies on omni-scale 2D convolutions with multi-scale context aggregation; CDNet uses combined depth space with 2D group convolutions; EfficientNet/MobileNetV2 use SE blocks with 2D global average pooling. A literal Conv2d→Conv1d substitution would either crash (BatchNorm2d receiving 3D inputs) or silently disable these central components. If the authors in fact converted all 2D ops (norm, pooling, attention), then the claim of 'simply substituting' is inaccurate, and there is no analysis showing the 1D versions preserve each method's intended mechanism. Either way, the baselines are likely systematically weakened, inflating DOLPHIN's claimed advantage. This is the single most load-bearing concern because the paper's central claim is empirical superiority over prior methods; if fairly adapted baselines close the gap, the contribution weakens substantially. A secondary consistency issue: Section V-E reports '30 times' repetitions, while Table III's caption states '50 times'; this must be reconciled for reproducibility.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DOLPHIN, a 1D convolutional model for online writer retrieval from Chinese handwritten phrases, combining temporal and frequency features through two new blocks (HFGA for gated cross-attention over high-frequency DWT sub-bands, and CAIR for channel activation), plus a Context-Aware FPN. It also introduces OLIWER, a large aggregation dataset built from CASIA-OLHWDB, DCOH-E, and SCUT-COUCH2009. Experiments on OLIWER, segmented CASIA-OLHWDB2, segmented DCOH-E, and SCUT-COUCH2009 report large improvements over adapted baselines, with ablations, efficiency comparisons, writer identification results, and cross-domain analyses on sampling frequency and pressure.","tokens_in":23693,"tokens_out":6309,"duration_ms":62928,"significance":"If the results hold, the paper makes two substantive contributions: a new large-scale benchmark (OLIWER) for a relatively underexplored task, and a strong retrieval model with clear ablations showing the value of each proposed component. The empirical work is extensive: four datasets, repeated trials with reported standard deviations, module ablations, efficiency measurements, and additional closed-set identification and cross-domain experiments. The promised release of code and dataset is a further practical strength. However, the headline claim of state-of-the-art performance depends on two load-bearing points that are not fully established: the fairness and completeness of the 2D-to-1D baseline adaptations, and the traceability of the OLIWER/DCOH-E construction, especially the provenance of 255 additional DCOH-E writers and the validation of the stroke-based phrase segmentation. The cross-domain conclusions are also mixed: DOLPHIN underperforms Sig2Vec in one transfer direction, which the authors acknowledge.","major_comments":[{"comment":"The central comparison rests on the claim that 2D baselines were converted to 1D 'by simply substituting the Conv2d module with the Conv1d module in Pytorch implementations.' This is not a well-defined operation for the strongest baselines: CAL uses counterfactual attention over 2D spatial feature maps, OSNet relies on omni-scale 2D convolutions with multi-scale context aggregation, CDNet uses 2D group convolutions in a combined depth space, and EfficientNet/MobileNetV2 use SE blocks with 2D global average pooling. A literal Conv2d-to-Conv1d substitution would either fail to run (e.g., BatchNorm2d with 3D inputs) or silently disable these central mechanisms. If the authors instead converted all associated 2D operations (normalization, pooling, attention), that conversion is not 'simple' and needs to be documented and analyzed. Without a specification of the exact mapping and ideally released code for the adapted baselines, the large margins in Table III may be inflated by systematically weakened baselines. Please provide the adapted implementations and a per-baseline description of how attention/pooling/SE mechanisms were mapped to 1D, and confirm that each baseline's core design was preserved.","section":"§V-E, Table III"},{"comment":"The construction of OLIWER is load-bearing for every reported comparison, and two parts are not adequately supported. First, Section IV states that DCOH-E is formed by adding 5,100 lines from 255 additional writers to the Chinese subset of DCOH, but it gives no source, citation, or collection details for these additional writers. Without provenance, the writer identities and sample counts in OLIWER and DCOH-E cannot be verified, and any corrupted identities would directly affect all retrieval labels. Second, the phrase segmentation in Algorithm 1 uses as threshold the time interval ranked at position N_Y+1; this heuristic is never validated against ground-truth character boundaries. Since CASIA-OLHWDB has character-level annotations, a comparison of Algorithm 1's output against those annotations would provide a quantitative check, and a manual or automated validation for DCOH-E should be reported. Please document the origin and licensing of the additional writer data, confirm that no writer overlaps across the three constituent datasets, and validate the segmentation heuristic.","section":"§IV, DCOH-E/OLIWER"}],"minor_comments":[{"comment":"There is an inconsistency in the reported number of repetitions: Section V-E says 'we repeated each experiment 30 times,' while the Table III caption says 'Each experiment is repeated 50 times.' Please correct this discrepancy and ensure the stated number matches the actual experiments.","section":"§V-E vs Table III caption"},{"comment":"The attention computation is written as w = k^T @ q, with k in R^{⌊L/2⌋×d} and q in R^{L×d}. With @ denoting matrix multiplication, k^T @ q is undefined because the inner dimensions do not match; the intended expression is presumably q @ k^T (or an equivalent transposed formulation), and the softmax dimension should be described consistently with that expression. Please correct the equation to make the HFGA block reproducible.","section":"Eq. (2)-(3), §III-B"},{"comment":"The cross-domain results are mixed: DOLPHIN underperforms Sig2Vec on OLHWDB2→DCOH-E (57.02% vs 62.28% Rank-1; 21.26% vs 24.55% mAP), while outperforming others in the reverse direction. This is acknowledged in the text, but the conclusion and abstract should be worded to avoid implying uniform superiority in cross-domain settings; a short discussion of why the proposed model is less transferable in that direction would be helpful.","section":"§V-I and Conclusion"},{"comment":"The ablation table is difficult to parse: several rows share the same combination of checkmarks in the main text (e.g., the rows referenced as 'line 5' and 'line 7' are not unambiguously labeled), and the textual references to line numbers do not align with a clear row enumeration. Please add explicit row labels or a legend so that each configuration can be identified.","section":"Table IV"},{"comment":"The text says the segmentation details are 'included in Procedure 1 in supplementary files,' but Algorithm 1 appears in the main text immediately afterward. Please reconcile this cross-reference so readers know where the algorithm is defined.","section":"§IV, DCOH-E bullet"}],"recommendation":"major_revision","confidential_remarks":"The two major comments are fixable but essential: the 2D-to-1D baseline conversion must be made reproducible and credible, and the OLIWER/DCOH-E provenance and segmentation validation must be documented. Given the paper's promise to release code and data, the editor should ask for the actual baseline adaptation code and dataset construction scripts during the revision. I do not see evidence of intentional misreporting; the repetition-count discrepancy is likely a typo, but it should be corrected. The paper is within scope for TIFS."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a real contribution to a small field. The OLIWER benchmark—674k phrase-level online Chinese samples from 1,731 writers—is the first large-scale public resource for online writer retrieval, and DOLPHIN is a sensible, carefully ablated baseline. Code and dataset are promised. That alone merits attention.\n\nWhat is new: DOLPHIN combines DWT-based high-frequency gated cross-attention (HFGA), CAIR channel-activation blocks, and Context-Aware FPN. The components are each known in spirit, but the combination for 1D handwriting sequences is new. The ablations show each piece buys something, and the identification experiments plus efficiency numbers round out the story. The related-work coverage is adequate.\n\nSoft spots, in order of importance:\n1. The baseline comparison. The paper says 2D models were adapted by 'simply substituting the Conv2d module with the Conv1d module.' For CAL, OSNet, and CDNet, the attention and pooling are 2D-specific; a literal substitution would either crash or silently disable those mechanisms. If the authors actually converted all the surrounding 2D ops, that is more than 'simply substituting,' and if they did not, the strong baselines are weakened. Either way, the reported margins over CAL et al. may be inflated. This is the load-bearing issue and needs a direct response.\n2. The 30 vs 50 repetition inconsistency. Section V-E says 30 repetitions; Table III caption says 50. Minor to fix but must be reconciled.\n3. DCOH-E provenance. 5,100 lines from 255 'additional writers' are added to the Chinese DCOH subset with no source cited. Since OLIWER is meant to be traceable, this should be documented.\n\nThe cross-domain analyses are post hoc but clearly labeled as exploratory; not a flaw.\n\nOverall: central claim is plausible but not fully established until the baseline adaptation is clarified. I would send this to peer review. The right reviewer will ask for revised baselines and provenance, but the benchmark and code are worth refereeing. For someone working on handwriting forensics or online writer retrieval, this is a useful resource.","headline":"A genuinely useful benchmark and strong baseline for online writer retrieval, but the head-to-head gains need a careful look because the 2D-to-1D baseline adaptations may have weakened the competition.","tokens_in":24252,"tokens_out":2106,"would_cite":true,"duration_ms":20231,"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":"This paper claims that online writer retrieval on short Chinese handwritten phrases improves sharply when the model learns temporal and frequency representations together, and it introduces the DOLPHIN model and the OLIWER dataset to back…","keywords":["online writer retrieval","handwritten phrase retrieval","temporal-frequency representation","discrete wavelet transform","gated cross-attention","channel activation","OLIWER dataset","Chinese handwriting"],"falsifier":"Inspect the 5,100 added DCOH-E lines: if the claimed 255 additional writers do not appear in the cited DCOH release or any named source, and if Algorithm 1's stroke-interval threshold, when compared against manual character boundaries on a random sample, frequently assigns strokes to wrong characters, then the OLIWER labels and every comparison built on them would be corrupted.","tokens_in":23133,"feed_emoji":"✍️","tokens_out":6435,"duration_ms":54419,"temperature":0.7,"pith_summary":"This paper tries to establish that online writer retrieval—finding all handwriting samples by the same person from short online handwritten phrases—can be made substantially more accurate by jointly learning temporal and frequency representations, and that the field has lacked a large-scale benchmark to measure this. To that end, it introduces DOLPHIN, a 1D CNN retrieval model, and OLIWER, a dataset of 674,017 Chinese handwritten phrases from 1,731 writers, and reports that DOLPHIN improves Rank-1 accuracy to 96.40% and mAP to 69.83% on OLIWER, against 93.32% and 61.01% for the best prior method. The paper also claims that cross-domain retrieval performance depends critically on aligning point sampling frequency and pressure information between datasets. A sympathetic reader would care because reliable phrase-level writer retrieval would support forensic search and handwriting analysis on the short, dynamic samples that dominate real online writing.","feed_headline":"A new model retrieves the right handwriting writer 96.4% of the time","feed_subtitle":"With a new 674,017-phrase benchmark, the model beats prior methods and shows that sampling rate and pressure matter.","key_machinery":"The load-bearing object is the HFGA block, a gated cross-attention module that applies a 1D discrete wavelet transform to split an input handwriting sequence into high-frequency and low-frequency sub-bands, then uses the high-frequency band as keys and values and the original temporal sequence as queries, with a tanh-gated residual connection. It lets the model amplify stroke-level details that are hard to see in the time domain alone. The CAIR block, a channel-split inverted residual with squeeze-excitation and channel shuffle, is the temporal counterpart that reduces channel redundancy, and a Context-Aware FPN adds multi-scale feature fusion with global context. Together these modules form DOLPHIN, a 1D CNN trained with Circle loss, OIM loss, and label-smoothed writer-ID loss, and evaluated by cosine-similarity ranking.","core_discovery":"The central claim is that a representation which combines the original temporal trace of an online handwriting sample with its high-frequency sub-bands, extracted via a discrete wavelet transform, captures more of a writer's individual style than temporal-only or existing retrieval representations. DOLPHIN operationalizes this with the HFGA block, which runs gated cross-attention between the temporal sequence and the high-frequency components so that local writing details such as stroke curvature and pressure variation are amplified, while the CAIR block reduces channel redundancy in the temporal backbone. Reports show DOLPHIN outperforming all compared methods on four datasets (OLIWER, segmented CASIA-OLHWDB2, segmented DCOH-E, and SCUT-COUCH2009), with the largest gains on mAP; ablation studies attribute the gains to CAIR, HFGA, and Context-Aware FPN, and cross-domain experiments support the claim that raising sampling frequency from 30Hz to 120Hz and adding pressure information narrows the distribution gap between handwriting domains.","pith_inferences":["The OLIWER dataset is an aggregation of existing sources rather than a newly collected corpus, so its utility depends on the integrity of writer identity and phrase segmentation in the source datasets; an independent audit of that lineage would be the strongest validation of the benchmark.","The conclusion that sampling frequency and pressure are the dominant domain-gap factors is supported by controlled experiments, but the direction of causality could be tested further by training on native 120Hz data with pressure artificially removed without resampling artifacts.","The architecture is language-agnostic in principle, but this paper only tests Chinese data; applying the same pipeline to English, Arabic, or Vietnamese online handwriting after phrase-level segmentation would test whether the temporal-frequency synergy transfers."],"forward_implications":["DOLPHIN sets a new state of the art for open-set online writer retrieval on phrase-level Chinese handwriting, including on mAP, where it surpasses the best prior method by 8.82 points on OLIWER.","OLIWER provides a community benchmark of 674,017 phrases from 1,731 writers, filling the large-scale dataset gap the paper identifies in online writer retrieval.","Because all models are evaluated with the same 14 hand-crafted time functions, the reported margins are attributable to architecture rather than input preprocessing.","Cross-domain results imply that future online handwriting datasets should record pressure and use a 120Hz sampling rate to make retrieval systems transfer across domains.","DOLPHIN's 2.14M parameters and roughly 9.95ms per-sample inference time make it plausible for real-time forensic search on large galleries."],"supporting_citations":[{"why":"Supplies the 52,220 text lines from 1,019 writers that become most of OLIWER's phrase samples after segmentation.","marker":"[13]"},{"why":"Supplies the 87,759 lines for DCOH-E, including pressure and timestamp features used in cross-domain experiments.","marker":"[14]"},{"why":"Supplies 145,000 phrase samples from 145 writers, the third source dataset aggregated into OLIWER.","marker":"[15]"},{"why":"Provides the gated cross-attention mechanism adopted in the HFGA block.","marker":"[7]"},{"why":"Provides the Inverted Residual structure that CAIR modifies and also serves as a baseline model.","marker":"[8]"},{"why":"Supplies the Selective Pooling scheme and an online-signature baseline that is the strongest comparator on two datasets.","marker":"[17]"},{"why":"Is the best prior method on OLIWER, so the Rank-1 and mAP comparisons anchor the claimed improvement.","marker":"[19]"},{"why":"Represents the offline writer retrieval baseline migrated for comparison.","marker":"[26]"},{"why":"Supplies the Global Context block fused into the Context-Aware FPN.","marker":"[12]"}],"fun_headline_variants":["DOLPHIN's temporal-frequency fusion outranks existing writer retrieval","Sampling rate and pressure are key to matching handwriting writers","New benchmark with 674K phrases sharpens online writer retrieval","High-frequency stroke features boost handwriting writer identification","Wavelet sub-bands and channel attention refine writer retrieval in DOLPHIN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that OLIWER is a valid collection of correctly labeled handwriting from 1,731 distinct writers, which requires that the 255 additional writers added to DCOH to form DCOH-E come from a traceable source and that Algorithm 1's stroke-interval threshold cuts text lines at true character boundaries.","fun_headline_variants_meta":{"raw":{"variants":["DOLPHIN's temporal-frequency fusion outranks existing writer retrieval","Sampling rate and pressure are key to matching handwriting writers","New benchmark with 674K phrases sharpens online writer retrieval","High-frequency stroke features boost handwriting writer identification","Wavelet sub-bands and channel attention refine writer retrieval in DOLPHIN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3342,"prompt_tokens":1011,"completion_tokens":2331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":2244}},"tokens_in":627,"tokens_out":2331,"duration_ms":19212,"temperature":1.0,"reasoning_tokens":2244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:41:50.084928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the 5,100 added DCOH-E lines: if the claimed 255 additional writers do not appear in the cited DCOH release or any named source, and if Algorithm 1's stroke-interval threshold, when compared against manual character boundaries on a random sample, frequently assigns strokes to wrong characters, then the OLIWER labels and every comparison built on them would be corrupted.","supporting_citations":[{"cited_title":"Flamingo: A Visual Language Model for Few-Shot Learning,","cited_arxiv_id":null,"evidence_quote":"Provides the gated cross-attention mechanism adopted in the HFGA block."},{"cited_title":"CASIA Online and Offline Chinese Handwriting Databases,","cited_arxiv_id":null,"evidence_quote":"Supplies the 52,220 text lines from 1,019 writers that become most of OLIWER's phrase samples after segmentation."},{"cited_title":"EGO-LM: An Efficient, Generic, and Out-of-the-box Language Model for Hand- written Text Recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the 87,759 lines for DCOH-E, including pressure and timestamp features used in cross-domain experiments."},{"cited_title":"SCUT- COUCH2009—A Comprehensive Online Unconstrained Chinese Handwriting Database and Benchmark Evalua- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies 145,000 phrase samples from 145 writers, the third source dataset aggregated into OLIWER."},{"cited_title":"MobileNetV2: Inverted Residuals and Linear Bottlenecks,","cited_arxiv_id":null,"evidence_quote":"Provides the Inverted Residual structure that CAIR modifies and also serves as a baseline model."},{"cited_title":"SynSig2Vec: Forgery-Free Learning of Dynamic Signature Represen- tations by Sigma Lognormal-Based Synthesis and 1D CNN,","cited_arxiv_id":null,"evidence_quote":"Supplies the Selective Pooling scheme and an online-signature baseline that is the strongest comparator on two datasets."},{"cited_title":"Counterfactual At- tention Learning for Fine-Grained Visual Categorization and Re-Identification,","cited_arxiv_id":null,"evidence_quote":"Is the best prior method on OLIWER, so the Rank-1 and mAP comparisons anchor the claimed improvement."},{"cited_title":"Towards Writer Retrieval for Historical Datasets,","cited_arxiv_id":null,"evidence_quote":"Represents the offline writer retrieval baseline migrated for comparison."},{"cited_title":"GCNet: Non- Local Networks Meet Squeeze-Excitation Networks and Beyond,","cited_arxiv_id":null,"evidence_quote":"Supplies the Global Context block fused into the Context-Aware FPN."}],"review_version":1}