{"id":"4f41ce35-4409-4c3e-a2fe-676ec4753f18","arxiv_id":"2506.01806","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ridgeformer uses a ViT plus cross-attention stage and multi-similarity loss to outperform prior methods on cross-domain fingerprint matching benchmarks.","lead":"This paper presents Ridgeformer, a two-stage transformer-based method for matching contactless and contact-based fingerprints by first extracting global features and then using cross-attention for fine-grained alignment. The authors report state-of-the-art error rates on two public datasets, which could make touchless fingerprint authentication more reliable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HKPolyU superiority claim is undercut by Table 1 itself: ML Fusion has higher TAR@FAR=0.01 (94.40%) than Ridgeformer (89.34%), so 'outperforms all previously benchmarked models' holds only for EER, not for the reported TAR metric.","rationale":"The reader's formal 'weakest_assumption' is about hyperparameter tuning on evaluation splits. That is a legitimate risk, but it is speculative and would require external evidence to confirm. The sharper, internally checkable problem is the TAR@FAR=0.01 contradiction between ML Fusion and Ridgeformer in Table 1. The reader did flag this inconsistency in the rationale, but did not make it the primary load-bearing concern. My stress-test places it at the center because it directly contradicts the unqualified superiority claim without relying on assumptions about tuning. I still view the paper as conditionally acceptable: the architecture is plausible, the ablation shows monotonic gains from both stages, and the code is released, so the claims can be verified. The requirement should be that the paper either qualify the comparison or explain the operating-point choice. Hence UNCHANGED relative to the reader's CONDITIONAL verdict.","tokens_in":7563,"tokens_out":4295,"duration_ms":45428,"concrete_test":"Re-extract the full ROC/DET curves for the HKPolyU contactless-to-contact verification protocol using the released Ridgeformer code and independently recompute ML Fusion under the same preprocessing and segmentation. If ML Fusion's TAR@FAR=0.01 remains 94.40 while Ridgeformer's is 89.34, replace the 'outperforms all' claim with a metric-specific statement (lowest EER only); if the table has a formatting typo (e.g., '4.0794.40' should be parsed differently), correct the table and re-evaluate the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and §4.4 is that Ridgeformer outperforms all previously benchmarked models on HKPolyU. Table 1 lists ML Fusion [20] at EER 4.07% and TAR@FAR=0.01 94.40%, while Ridgeformer is 2.83% EER and 89.34% TAR@FAR=0.01. The EER is better, but at the operating point FAR=0.01 the baseline has a 5.06-point higher TAR. Since TAR@FAR is the more security-relevant metric and is reported in the same table, the unqualified 'outperforms all' claim is internally contradicted. The paper never discusses this trade-off; §4.4 only cites EER <3%. It is mathematically possible for one model to have lower EER and lower TAR at a fixed FAR when ROC curves cross, so this is not an impossibility, but it means 'best cross-domain verification accuracy' is not established without a threshold-preference argument. The conclusion should be qualified to 'lowest EER' or the operating point must be justified. This concern is load-bearing because the headline result depends on an unqualified comparison, and it can be settled directly from the table or its underlying scores.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Ridgeformer, a two-stage transformer architecture for cross-domain fingerprint recognition: Stage 1 extracts global features with a Vision Transformer, and Stage 2 performs fine-grained cross-sample alignment via a cross-attention module. Training combines a multi-similarity loss over three similarity matrices (CL2CL, CL2CB, CB2CB). The method is evaluated for 1:1 verification and 1:N identification on HKPolyU and RidgeBase, with an ablation study. The authors report state-of-the-art EER on HKPolyU CL2CB (2.83%) and competitive or better results on RidgeBase, and they release code.","tokens_in":7874,"tokens_out":2611,"duration_ms":27903,"significance":"If the claims hold, Ridgeformer would be a meaningful advance in contactless-to-contact fingerprint matching, a domain where domain shift between sensors is a known challenge. The paper has concrete strengths: it uses public benchmarks (HKPolyU, RidgeBase), reports comparisons including a COTS matcher (Verifinger), provides an ablation of the two stages, and makes code available. These features make the work reproducible in principle and of interest to the biometrics community. However, the headline claim of 'outperforms all previously benchmarked models' is not fully supported by the reported metrics, and there are open questions about the training/evaluation protocol. The central empirical contribution therefore needs revision rather than being acceptable as is.","major_comments":[{"comment":"The claim that 'Ridgeformer outperforms all previously benchmarked models on the HKPolyU dataset' is internally contradicted by Table 1. On HKPolyU CL2CB, Ridgeformer achieves EER 2.83% and TAR@FAR=0.01 of 89.34%, while ML Fusion [20] has EER 4.07% but TAR@FAR=0.01 of 94.40%. The ROC curves cross: Ridgeformer is better at the equal-error operating point but worse at the fixed FAR=0.01 operating point, which is the more security-relevant metric reported in the same table. Section 4.4 cites only EER, ignoring the TAR metric. The conclusion should be qualified to 'lowest EER among compared methods' or the paper must justify why the FAR=0.01 operating point is not decisive.","section":"Table 1 and Section 4.4"},{"comment":"The training/evaluation protocol for HKPolyU is ambiguous and potentially leaky. Section 4.2.2 states that evaluation uses the HKPolyU testing split, disjoint from training data, but Section 4.1 says 'During fine-tuning on the HKPolyU dataset [2]' the margin and threshold were changed. It is not specified whether this fine-tuning was performed on the training split only or on the full HKPolyU dataset. Furthermore, hyperparameters (alpha_pos=2.0, alpha_neg=40.0, margin, threshold tau, learning-rate decay epochs) are described as 'determined through experimentation' without stating whether the evaluation split was used for model selection. If the test split influenced fine-tuning or hyperparameter choice, the reported EER and TAR are optimistic. The authors must clarify the exact data flow.","section":"Section 4.1 and Section 4.2.2"},{"comment":"The claimed superiority on RidgeBase is not established because the comparison set is incomplete. For the CL2CB protocol, Table 1 lists only Verifinger and Ridgeformer; no learning-based baselines such as ML Fusion [20] or MANet [13] are reported. For the CL2CL protocol, only Verifinger and AdaCos(CNN) are compared. The abstract and Section 4.4 assert 'outperforms existing methods, including COTS solutions,' but without the relevant learned baselines on the same protocol, this is an unsupported generalization. The authors should add the missing baseline comparisons or restrict the claim to the methods actually evaluated.","section":"Table 1 and Table 2 (Ridgebase Task 1)"}],"minor_comments":[{"comment":"The title contains a typo: 'Mutli' should be 'Multi'.","section":"Title and Abstract"},{"comment":"The phrase 'an multi-stage training strategy' in the second contribution should be 'a multi-stage training strategy'.","section":"Section 1, Contributions"},{"comment":"The definitions of P and N in the multi-similarity loss are incomplete in the displayed equations: the subscript of S_ij is lost, making it hard to identify which pairwise similarity is being exponentiated. Please restate the equations in the standard form of [17].","section":"Section 3.3, Equation formatting"},{"comment":"The 100.00% Rank-1 and Rank-10 on RidgeBase CL2CL is remarkable and is not discussed. A sentence explaining why this is plausible (e.g., gallery size, high-quality images, near-perfect separation in the learned embedding) would help the reader assess the result.","section":"Table 2"},{"comment":"The statement about ISPFDv2's evaluation protocol is confusing: the paper says the official protocol requires a 'random' 50-50 split with three rounds of validation, then says the authors do not report results due to non-reproducibility. It would be clearer to state explicitly that ISPFDv2 was used only for training and not for evaluation.","section":"Section 4.2.1"},{"comment":"Reference [22] is the same paper as reference [1] (Dosovitskiy et al., 'An image is worth 16x16 words'). Use a single citation with the correct conference venue.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is worth reading if you work on contactless fingerprints. The two-stage design—global ViT embedding, then cross-attention between token sets for fine-grained alignment—is sensible, and the ablation shows each stage buys a real EER/TAR improvement. They also ship a code link and are upfront about why ISPFDv2 is not reported (non-reproducible random splits). That is honest.\n\nThe headline claim does not survive contact with Table 1. On HKPolyU, Ridgeformer gets EER 2.83% versus ML Fusion's 4.07%, but ML Fusion gets TAR@FAR=0.01 of 94.40% versus Ridgeformer's 89.34%. The abstract and §4.4 say 'outperforms all previously benchmarked models.' That is true for EER, false for the security-relevant operating point they report in the same table. ROC curves cross; either a threshold-preference argument or a qualified 'lowest EER' is needed. This is fixable in revision but it is a real overstatement.\n\nTwo other soft spots. First, the paper cites Grosz et al. [15] (minutiae-guided ViT embeddings) and MRA-GNN [16] but does not compare to either; that makes the 'first ViT for contactless-to-contact matching' claim shaky if [15] already does cross-domain work. Second, the hyperparameters are 'determined through experimentation' (§4.1) with no explicit statement that tuning was done on held-out validation, and there are no error bars. Those are standard concerns for this venue, not fatal.\n\nWhat holds up: the method itself is plausible, the ablation is clean, the Ridgebase numbers look strong, and the code release is a plus. The central architecture is not a gimmick.\n\nRecommendation: send to peer review, but insist the authors reconcile the TAR/EER conflict and either add the cited baselines or justify their omission. As is, the paper should be accepted only after that revision.","headline":"A solid two-stage ViT fingerprint matcher with a real but fixable overclaim: their own Table 1 contradicts the 'outperforms all' headline on TAR@FAR=0.01.","tokens_in":8371,"tokens_out":2519,"would_cite":false,"duration_ms":26505,"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":"Ridgeformer claims best-reported contactless-to-contact fingerprint verification on HKPolyU, with EER 2.83%.","keywords":["contactless fingerprint recognition","cross-domain fingerprint matching","vision transformer","cross-attention","multi-similarity loss","biometric verification","fingerprint identification","Ridgeformer"],"falsifier":"Re-run the HKPolyU contactless-to-contact protocol with all hyperparameters (including $\\alpha_{\\text{pos}}=2.0$, $\\alpha_{\\text{neg}}=40.0$, margin, $\\tau$, and the learning-rate schedule) chosen only on a held-out validation fold, never on the test split. If the resulting EER rises above 3% or fails to beat MANet's 4.13%, the paper's superiority claim is refuted.","tokens_in":7395,"feed_emoji":"🖐️","tokens_out":5608,"duration_ms":50615,"temperature":0.7,"pith_summary":"Ridgeformer sets out to solve cross-domain fingerprint matching, where a fingerprint photographed by a contactless sensor must be matched against a contact-based scan of the same finger. The authors propose a two-stage transformer pipeline: a Vision Transformer extracts global patch-level features, and a cross-attention stage aligns those patches between a probe and a gallery image before computing similarity. Training uses a multi-similarity loss applied to contactless-to-contactless, contactless-to-contact, and contact-to-contact similarity matrices. On HKPolyU contactless-to-contact verification, the method reports an equal error rate of 2.83% and a true-accept rate of 89.34% at FAR=0.01, which the authors state outperforms all previously benchmarked models on that dataset. The claim matters because reliable contactless-to-contact matching is the obstacle to hygienic, touch-free fingerprint authentication in real deployments.","feed_headline":"Ridgeformer cuts cross-domain fingerprint EER below 3%","feed_subtitle":"Two-stage transformer with cross-attention beats prior methods and COTS on contactless-to-contact matching.","key_machinery":"The machine is a two-stage pipeline named Ridgeformer. Stage 1 uses a Vision Transformer to split each fingerprint into $T$ patch tokens, pools them with global average pooling, and projects the pooled vector through an MLP to get a global embedding. Stage 2 takes the token sets of a probe and a gallery image, concatenates them, and passes them through a cross-attention transformer; the attended tokens are pooled and compared by cosine similarity. Training uses the multi-similarity loss with hard positive/negative mining, computing the loss on three similarity matrices: contactless-to-contactless, contactless-to-contact, and contact-to-contact. The named identity is the combination of the multi-stage architecture and the three-way contrastive objective.","core_discovery":"The central claim is that a hierarchical global-to-local transformer, trained with a contrastive multi-similarity loss across both same-domain and cross-domain pairs, yields the best reported contactless-to-contact fingerprint verification on HKPolyU, with EER 2.83% and TAR 89.34% at FAR=0.01, beating MANet (EER 4.13%) and a commercial COTS matcher (EER 19.31%). On the RidgeBase benchmark, the same model achieves EER 5.25% for contactless-to-contact and 7.60% for contactless-to-contactless verification, with perfect Rank-1 identification (100%) in the contactless-to-contactless setting. The authors attribute the improvement to Stage 2's cross-attention refinement: the ablation shows EER dropping from 3.74% with only Stage 1 to 2.83% after adding Stage 2 and fine-tuning.","pith_inferences":["A natural extension, not tested in the paper, is whether the same global-then-cross-attention recipe transfers to other cross-sensor biometrics (e.g., iris or face) where domain shift comes from sensor geometry and illumination rather than fingerprint ridge distortion.","Because the paper does not report validation-based hyperparameter selection, the fairest reading is that the 2.83% EER is an upper bound on what a practitioner would get with a clean tuning protocol; an independent replication with hyperparameters chosen on a disjoint validation split would settle this.","The near-perfect contactless-to-contactless identification on RidgeBase could also be a ceiling effect of the small evaluation set (25 subjects); testing on a larger cohort would reveal whether the 100% Rank-1 figure is meaningful or an artifact of gallery size.","The paper's decision to omit ISPFDv2 results due to protocol non-reproducibility suggests that standardized evaluation protocols for contactless fingerprint benchmarks are a prerequisite for comparing methods; a shared protocol with fixed train/test splits would make future claims directly comparable."],"forward_implications":["If the reported numbers hold, contactless-to-contact verification is no longer the weak link: a touchless capture can be matched to legacy contact-scan databases at an EER below 3% on HKPolyU.","The ablation shows each stage contributes: global features alone give EER 3.74%, adding cross-attention lowers it to 3.04%, and fine-tuning on the target dataset brings it to 2.83% with TAR rising from 84.16% to 89.34%.","On RidgeBase, the same model reaches 100% Rank-1 identification for contactless-to-contactless matching, suggesting the learned embedding separates identities cleanly in that protocol.","The method's advantage over COTS (EER 19.31% vs 2.83% on HKPolyU) implies that learned cross-domain embeddings can replace or supplement commercial matchers in contactless enrollment scenarios."],"supporting_citations":[{"why":"Provides the HKPolyU benchmark, the RTPS+DCM baseline, and part of the training corpus.","marker":"[2]"},{"why":"Provides the RidgeBase dataset and the Task 1 evaluation protocol used for the second benchmark.","marker":"[3]"},{"why":"MANet is the strongest prior baseline that Ridgeformer reports beating on HKPolyU.","marker":"[13]"},{"why":"Multi-Siamese is an earlier deep-learning baseline for contactless-to-contact matching.","marker":"[14]"},{"why":"Supplies the multi-similarity loss and the hard positive/negative mining procedure used in training.","marker":"[17]"},{"why":"ML Fusion provides the AdaCos+CNN baseline that Ridgeformer compares against on both datasets.","marker":"[20]"},{"why":"The Vision Transformer architecture is the Stage 1 backbone that produces patch tokens and global embeddings.","marker":"[22]"}],"fun_headline_variants":["Ridgeformer hits 2.83% EER on contactless fingerprint matching","Cross-domain fingerprint EER below 3% with two-stage transformer","Two-stage transformer beats COTS on contactless fingerprint recognition","Hierarchical transformer achieves best fingerprint EER 2.83%","Global-to-local transformer tops fingerprint matching, EER 2.83%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported superiority rests on hyperparameters that the authors say were 'determined through experimentation,' with no statement that this tuning was done on held-out validation data rather than the evaluation splits.","fun_headline_variants_meta":{"raw":{"variants":["Ridgeformer hits 2.83% EER on contactless fingerprint matching","Cross-domain fingerprint EER below 3% with two-stage transformer","Two-stage transformer beats COTS on contactless fingerprint recognition","Hierarchical transformer achieves best fingerprint EER 2.83%","Global-to-local transformer tops fingerprint matching, EER 2.83%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1336,"prompt_tokens":898,"completion_tokens":438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":344}},"tokens_in":514,"tokens_out":438,"duration_ms":5100,"temperature":1.0,"reasoning_tokens":344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:32:20.121619+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the HKPolyU contactless-to-contact protocol with all hyperparameters (including $\\alpha_{\\text{pos}}=2.0$, $\\alpha_{\\text{neg}}=40.0$, margin, $\\tau$, and the learning-rate schedule) chosen only on a held-out validation fold, never on the test split. If the resulting EER rises above 3% or fails to beat MANet's 4.13%, the paper's superiority claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HKPolyU benchmark, the RTPS+DCM baseline, and part of the training corpus."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RidgeBase dataset and the Task 1 evaluation protocol used for the second benchmark."},{"cited_title":"Matching contactless and contact- based conventional fingerprint images for biometrics identification,","cited_arxiv_id":null,"evidence_quote":"MANet is the strongest prior baseline that Ridgeformer reports beating on HKPolyU."},{"cited_title":"Ridgebase: A cross-sensor multi- finger contactless fingerprint dataset,","cited_arxiv_id":null,"evidence_quote":"Multi-Siamese is an earlier deep-learning baseline for contactless-to-contact matching."},{"cited_title":"Syncolfinger: Synthetic contact- less fingerprint generator,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-similarity loss and the hard positive/negative mining procedure used in training."},{"cited_title":"Multi loss fusion for matching smartphone captured contactless finger images,","cited_arxiv_id":null,"evidence_quote":"ML Fusion provides the AdaCos+CNN baseline that Ridgeformer compares against on both datasets."},{"cited_title":"Towards more accurate contact- less fingerprint minutiae extraction and pose-invariant matching,","cited_arxiv_id":null,"evidence_quote":"The Vision Transformer architecture is the Stage 1 backbone that produces patch tokens and global embeddings."}],"review_version":1}