acceptodds
Under review as a conference paper at ICLR 2027

SigDINO: Adapting DINOv3 with Local Structural Matching for Writer-Independent Offline Signature Verification

Abstract

Offline handwritten signature verification aims to determine the authenticity of a signature from static signature images alone. Existing methods often rely on large amounts of training data to learn discriminative signature representations. However, when only limited training data are available, their performance tends to become unstable and their generalization ability often degrades in cross-language evaluation. To address these issues, we propose SigDINO, a writer-independent offline signature verification method that adapts DINOv3, a large scale self-supervised vision model with transferable visual representations, to the signature domain. Specifically, we perform self-supervised adaptation of the pretrained model using synthetic signature images, allowing it to better capture the sparse stroke distribution of offline signatures. Intermediate stroke responsive patch tokens from the adapted transformer serve as the main local representation, while cross layer residual information from deeper layers further enhances them. These representations are then compared by the local structural matching module to compute the verification distance between signature pairs. Extensive experiments on four benchmark datasets show that SigDINO achieves state-of-the-art or competitive performance, while maintaining stable verification under limited writer settings and strong generalization in cross-language evaluation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
offline signature verification, writer-independent verification, self-supervised learning, DINOv3, local structural matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_s2gTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SigDINO, and extention of DINO for offline signature verification. The main difference between the original is that the proposed method extracts intermediate layer patch features to preserve local details, then uses structure matching to establish correspondences between patches. It then uses the matching costs to determine the distance which is used for signature verification. They test on four datasets including cross-language evaluations.

Critical strengths

The authors have strong empirical performance across various languages. In addition, the results from DetailSemNet supports the base idea of this paper that local patches/details distance can improve signature verification performance. Even though the paper is incremental, there are some interesting parts.

Critical weaknesses
  1. It's not totally discounting, but most of the performance seems to come from the DINOv3. For example, in Table 6, raw DINOv3 beats 7/11, 10/11, 4/5, 1/1 of the competitors respectively without the proposed method. Of course, I can understand that the last little bit is important, it just weakens the claims by SigDINO.

  2. Equation (5) is very tuned. Specifically, layer 17 and the residual connected layer 23 are used. This seems hyper specific to this case and might not generalize to other problem sets. The authors do have some ablation, in Table 6, they report a few other stroke token layer patch proposals. However, I would like to see a larger range of tests, especially because it seems strange that the performance increases between 14 and 20. In addition, the choice of layer 23 isn't looked at all. A bigger problem is that it's not clear if the choice of layer number was chosen with a validation set or with the test set.

Questions for authors
  1. Why is ChiSig left out of some of the evaluations (ex:cross language and the appendix)?

  2. How are any of the specific parameters selected (ex: weight sensitivity vs matching distance, layer number, patch size, etc.)? Specifically, I know there are tables that show the empirical results, but are these empirical results done using a validation set? or are they chosen based on the test set?

3a. In Eq. (5), the equation is hi=p~17i+(p~17i−p~23i)h^i=\tilde{p}^i_{17}+(\tilde{p}^i_{17}-\tilde{p}^i_{23}). Isn't this just hi=2p~17i−p~23ih^i=2\tilde{p}^i_{17}-\tilde{p}^i_{23}? what's the rationale for emphasizing L17 by double and subtracting L23?

3b. I don't think Figure 3 accurately reflects the Eq. (5). According to the caption, "Residual" is the difference between L17 and L23. But, looking at L17 and L23, I don7t think that's true. If you look closely, there are parts not lit up by theither L17 or L23 but are lit up in "Residual". If it was the difference, these areas should also not be lit up (ex: the areas between characters). In addition, there are places lit up by both and are still lit up on "Residual" (ex: the centers of characters). In addition "Enchanced" should be the sum of "Residual" and L17, but it doesn't appear to be the case at all. For example the upper right corner is lit up on "Residual" but not on "Enhanced". Actually, "Enhanced" looks almost the same as L17 except in a different color pallet.

8 comments
AuthorsOpenReview ↗

Response to Reviewer s2gT: W1- Novelty and isolation of component contributions

Response to Reviewer s2gT: W1 - Novelty and isolation of component contributions

We sincerely thank you for these thoughtful comments and for reading Table 6 so closely. Few prior works have explored large-scale self-supervised models for offline signature verification. This requires bridging the distribution gap between natural images and signature images, where discriminative cues lie in local stroke structures rather than object-level semantics. Meanwhile, we would like to clarify why pre-trained representations alone cannot tackle the two challenges we focus on: limited training writers in downstream tasks and cross-lingual generalization, and we provide new experimental evidence for each of our two core designs.

1. Signature-domain adaptation for limited-writer training.

Signature-domain adaptation uses unlabeled synthetic signatures prior to downstream supervised training, with the aim of reducing reliance on a large set of labeled training writers.

We realize that one point in our ablation tables was not stated clearly. "Raw DINOv3" in Tables 6 and 7 retains our cross-layer feature processing and local structural matching module; only the signature-domain self-supervised adaptation is removed. Its numbers therefore reflect unadapted DINOv3 combined with our local structural matching, not raw DINOv3 global features. We will relabel this row in the revision.

Table R1. Effect of signature-domain adaptation under five training writers. Both settings use the same verification architecture. Reproduced from Table 7; EER (%), lower is better.

Backbone settingBengaliHindiCEDARChiSigAverage
Without signature-domain adaptation20.9211.5117.5816.3816.60
With signature-domain adaptation5.859.1810.5510.629.05

With the verification architecture held fixed, signature-domain adaptation lowers EER on all four datasets, reducing the average from 16.60% to 9.05% (-7.55 points). This provides direct evidence of its benefit under limited-writer supervision.

2. Local structural matching for cross-language transfer.

Local structural matching addresses the need to compare local stroke structures across writing systems. Aggregating each signature independently into a single vector before comparison can weaken subtle local stroke differences. SigDINO therefore preserves foreground patch representations, establishes bidirectional soft correspondences, and then measures local discrepancies and aggregates the matched relations, so that the resulting representation reflects interactions between the two signatures rather than each one in isolation.

To examine this design directly, we ran a head-replacement experiment. We kept the signature-adapted backbone, the cross-layer enhancement, and the foreground feature extractor unchanged, and replaced only the local structural matching module with an MLP verification head of comparable parameter count; verification then uses the cosine distance between independently computed signature embeddings.

Table R2. Verification-head comparison; EER (%), lower is better. Rows give the training dataset and head; columns give the test dataset. The "Local structural matching" rows are from Tables 1, 2 and 4.

Training setHeadBengaliHindiCEDAR
BengaliMLP2.1722.3910.58
BengaliLocal structural matching (ours)1.0311.755.87
HindiMLP10.384.9214.57
HindiLocal structural matching (ours)3.261.937.75
CEDARMLP13.3619.381.23
CEDARLocal structural matching (ours)7.7510.160.07

The MLP variant reaches an average in-domain EER of 2.77% over the three datasets and 15.11% over the six cross-language transfer directions; the corresponding figures for local structural matching are 1.01% and 7.76%, with lower EER in every one of the six transfer directions. The gap is modest in domain and much larger across languages, which is consistent with our motivation: preserving and comparing local correspondences before final aggregation matters most when the two signatures come from different writing systems.

3. Positioning of the contribution.

We would summarize our contribution as combining signature-domain representation adaptation with foreground local structural matching to address limitations that general-purpose pretrained representations exhibit on signature images: unlabeled signature adaptation supports training with few writers, while bidirectional local correspondences and matched-relation aggregation support cross-language verification. The ablations above indicate that the original pretrained features alone are not sufficient for this task, and that the two proposed components contribute substantial additional gains on top of them.

AuthorsOpenReview ↗

Response to Reviewer s2gT: W2/Q3a - Layer selection and the residual formulation in Equation (5)

Response to Reviewer s2gT: W2/Q3a - Layer selection and the residual formulation in Equation (5) Thank you for this valuable question. Equation (5) is algebraically equivalent to 2p~i17−p~i232\tilde{p}_i^{17}-\tilde{p}_i^{23}. We adopt this residual formulation to clarify our design rationale: we build upon the representation at layer L17, and inject the difference between L17 and the final layer L23 as a corrective signal.

Why intermediate layers? Offline signature verification relies primarily on stroke geometry — spatial positions, connectivity and local layout — rather than on object-level semantic concepts. Analyses of pretrained visual backbone representations indicate that shallow layers retain texture detail, intermediate layers capture structural layout, and deep layers move toward abstract semantics. Intermediate layers therefore offer a favourable balance between spatial awareness and contextual structure, which benefits position-sensitive tasks (Amir et al., 2021). DINOv3 shows the same tendency: in its per-layer analysis, geometry-focused tasks perform better below the final layer, and its dense prediction pipeline aggregates several intermediate-layer features. These findings motivate the use of intermediate-layer features, while the exact indices are determined by ablation.

Why L17 and L23? Our ViT-L/16 backbone contains 24 Transformer blocks with zero-based indexing. L17 lies in the upper-middle range and L23 corresponds to the final block. Table R1 reports the ablation over alternative layer pairs and fusion strategies.

Table R1. Ablation on layer combinations and fusion strategies (EER, %). ChiSig is evaluated under the 1,658-writer protocol.

Feature configurationBengaliHindiCEDARChiSigAvg
Single layer
L171.652.380.141.671.46
L232.933.470.224.622.81
Simple fusion of L17 and L23
L17 + L232.042.080.143.681.99
L17 − L231.372.310.221.981.47
concat(L17, L23)2.312.220.431.711.67
Residual with other layer pairs
L14 + (L14 − L17)1.081.960.071.881.25
L14 + (L14 − L20)1.621.960.221.961.44
L14 + (L14 − L23)1.121.940.142.061.31
L17 + (L17 − L20)1.752.050.221.791.45
L20 + (L20 − L23)1.432.330.293.041.77
Ours
L17 + (L17 − L23)1.031.930.071.641.17

Our configuration achieves the lowest average EER on all four datasets. Using L17 alone gives 1.46%, and simple fusions of L17 and L23 — addition, subtraction, concatenation — all remain at or above that level, so the gain does not come from merely having access to L23. Adopting the residual form lowers the average across all layer pairs tested, and among them L17 with a correction from L23 is the strongest: holding the base at L17, taking the correction from L23 rather than L20 reduces the average from 1.45% to 1.17%. This supports the design: L17 is retained as the primary local representation, and a cross-layer correction derived from a deeper block enhances it.

Revised Figure 3. To further verify and visually demonstrate our design, we have updated Figure 3 in the main text to an occlusion-sensitivity map. We mask each input patch sequentially and record the variation of matching distance; locations with larger distance shifts are highlighted with higher heat response. This visualization intuitively illustrates how we retain L17 as the core local representation and leverage L23 to produce a cross-layer correction term for representation enhancement.

AuthorsOpenReview ↗

Response to Reviewer s2gT: Q1 - ChiSig in cross-language and appendix evaluations

Response to Reviewer s2gT: Q1 - ChiSig in cross-language and appendix evaluations

Thank you for this suggestion. We are happy to add cross-language evaluation on ChiSig, which provides a strong test of transfer across writing systems: Chinese signatures differ substantially from those in the Bengali, Hindi and CEDAR datasets in character structure and stroke arrangement. Few studies in the public literature report cross-language signature verification involving ChiSig, which is the main reason the original manuscript did not include cross-language baseline comparisons for this dataset; we will clarify this in the revision.

Of the 1,658 eligible ChiSig writers, only 400 have both genuine and forged signatures; the remaining 1,258 (75.87%) have genuine signatures only. We therefore evaluated two protocols. The 395-writer balanced protocol uses 395 writers with exactly five genuine and five forged signatures each, under writer-disjoint train/test splits. The full 1,658-writer protocol retains all eligible writers under a writer-disjoint 829/829 split, which reflects the practical setting where forged signatures are scarce.

Table R1. In-domain and cross-language EER (%) under the two ChiSig protocols; lower is better. Rows are test datasets, columns are training datasets. A single value means the result is identical under both protocols; "395 / 1658" gives each protocol's result.

Test / TrainBengaliHindiCEDARChiSig
Bengali1.033.267.7513.19 / 10.78
Hindi11.751.9310.1614.21 / 10.19
CEDAR5.877.750.0716.45 / 6.74
ChiSig12.13 / 20.4813.15 / 21.6916.19 / 20.520.61 / 1.64

Under the 395 protocol the average EER over the six cross-language directions involving ChiSig is 14.22%. Given the substantial differences in character structure and stroke organization between Chinese and the other three writing systems, this indicates reasonable cross-writing-system transfer.

Training on ChiSig and testing on the other three datasets yields averages of 14.62% (395) and 9.24% (1658), so broader Chinese training coverage improves transfer outward. The reverse direction behaves differently: training on the other three and testing on ChiSig gives 13.82% (395) and 20.90% (1658). Under the 1658 protocol the Chinese test set spans a more diverse writer population, and the resulting shift in the distance distribution moves the operating point, raising the EER. Transfer to this broader Chinese distribution therefore still has room for improvement.

The two protocols are complementary: the 395 protocol characterises cross-language performance under class balance, while the 1658 protocol reflects the practical case where forgeries are scarce. We will report both protocols in the revision to give a fuller picture of the model's performance.

Reviewer_33W4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel writer-independent offline signature verification framework based on DINOv3, termed SigDINO. The method first performs self-supervised adaptation of DINOv3 to the signature domain using paired signature images. Subsequently, patch tokens from an intermediate layer of the adapted model are used as local stroke representations. These representations are enhanced with cross-layer residual information, while foreground stroke regions are retained for matching. Finally, a bidirectional local structural matching module compares signature pairs and determines their authenticity by combining token-level correspondence distance with a learned matching-embedding distance. Experiments are conducted on four public datasets, covering standard writer-independent verification and cross-language transfer settings. The results show that the proposed method achieves competitive or state-of-the-art performance. Ablation studies further analyze the contributions of the designed components.

Critical strengths
  1. Comprehensive experimental evaluation. SigDINO is evaluated on four public datasets, in cross-language transfer settings, and under a challenging five-writer training protocol. Multiple ablation studies further analyze the effects of the backbone, domain adaptation, feature-layer selection, and local matching components.
  2. Clear overall organization. The paper is organized around signature-domain self-supervised adaptation, hierarchical local feature extraction, and local structural matching, with a clear correspondence between the methodological design and the experimental analysis.
Critical weaknesses

Altough the proposed method has been validated on multiple datasets, I believe that several issues still require further investigation and verification:

  1. The ablation results indicate that most of the performance improvement comes from the DINOv3 backbone, while the additional gains brought by the proposed domain adaptation and local structural matching designs are relatively limited.
  2. The model-selection protocol is unclear. The paper does not specify whether an independent validation set is used or whether hyperparameters, including the feature layer, SSL configuration, and distance weights, are selected based on test-set performance.
  3. The statistical reliability is insufficiently evaluated. CEDAR contains only five test writers, but the paper does not report repeated splits, confidence intervals, or statistical significance analysis. The five-writer training experiments also do not clarify whether multiple random writer selections are performed.
  4. The generalization evaluation remains limited. The cross-language experiments do not include ChiSig, and the method is not evaluated under domain shifts caused by different acquisition devices, image quality, or real document environments.
  5. A complete computational-cost analysis is missing. The method uses high-resolution inputs, a DINOv3 ViT-L/16 backbone, and bidirectional token matching, which may introduce considerable computational overhead for this task. However, the paper does not report the parameter count, FLOPs, GPU memory consumption, or inference time.
Questions for authors

Could the authors clarify whether an independent validation set was used to select the feature layers, SSL configuration, and distance weights? Please also report the computational cost, broader cross-domain evaluation, and performance when downstream training uses only genuine signatures without forgery samples.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 33W4: W1 - Contribution of the backbone and the proposed components

Response to Reviewer 33W4: W1 - Contribution of the backbone and the proposed components

We sincerely thank you for this comment on how much of the improvement comes from the DINOv3 backbone.

Few prior works have explored large-scale self-supervised models for offline signature verification. This requires bridging the distribution gap between natural images and signature images, where discriminative cues lie in local stroke structures rather than object-level semantics. Meanwhile, we would like to clarify why pre-trained representations alone cannot tackle the two challenges we focus on: limited training writers in downstream tasks and cross-lingual generalization, and we provide new experimental evidence for each of our two core designs.

1. Signature-domain adaptation for limited-writer training.

Signature-domain adaptation uses unlabeled synthetic signatures prior to downstream supervised training, with the aim of reducing reliance on a large set of labeled training writers.

We realize that one point in our ablation tables was not stated clearly. "Raw DINOv3" in Tables 6 and 7 retains our cross-layer feature processing and local structural matching module; only the signature-domain self-supervised adaptation is removed. Its numbers therefore reflect unadapted DINOv3 combined with our local structural matching, not raw DINOv3 global features. We will relabel this row in the revision.

Table R1. Effect of signature-domain adaptation under five training writers. Both settings use the same verification architecture. Reproduced from Table 7; EER (%), lower is better.

Backbone settingBengaliHindiCEDARChiSigAverage
Without signature-domain adaptation20.9211.5117.5816.3816.60
With signature-domain adaptation5.859.1810.5510.629.05

With the verification architecture held fixed, signature-domain adaptation lowers EER on all four datasets, reducing the average from 16.60% to 9.05% (-7.55 points). This provides direct evidence of its benefit under limited-writer supervision.

2. Local structural matching for cross-language transfer.

Local structural matching addresses the need to compare local stroke structures across writing systems. Aggregating each signature independently into a single vector before comparison can weaken subtle local stroke differences. SigDINO therefore preserves foreground patch representations, establishes bidirectional soft correspondences, and then measures local discrepancies and aggregates the matched relations, so that the resulting representation reflects interactions between the two signatures rather than each one in isolation.

To examine this design directly, we ran a head-replacement experiment. We kept the signature-adapted backbone, the cross-layer enhancement, and the foreground feature extractor unchanged, and replaced only the local structural matching module with an MLP verification head of comparable parameter count; verification then uses the cosine distance between independently computed signature embeddings.

Table R2. Verification-head comparison; EER (%), lower is better. Rows give the training dataset and head; columns give the test dataset. The "Local structural matching" rows are from Tables 1, 2 and 4.

Training setHeadBengaliHindiCEDAR
BengaliMLP2.1722.3910.58
BengaliLocal structural matching (ours)1.0311.755.87
HindiMLP10.384.9214.57
HindiLocal structural matching (ours)3.261.937.75
CEDARMLP13.3619.381.23
CEDARLocal structural matching (ours)7.7510.160.07

The MLP variant reaches an average in-domain EER of 2.77% over the three datasets and 15.11% over the six cross-language transfer directions; the corresponding figures for local structural matching are 1.01% and 7.76%, with lower EER in every one of the six transfer directions. The gap is modest in domain and much larger across languages, which is consistent with our motivation: preserving and comparing local correspondences before final aggregation matters most when the two signatures come from different writing systems.

3. Positioning of the contribution.

We would summarize our contribution as combining signature-domain representation adaptation with foreground local structural matching to address limitations that general-purpose pretrained representations exhibit on signature images: unlabeled signature adaptation supports training with few writers, while bidirectional local correspondences and matched-relation aggregation support cross-language verification. The ablations above indicate that the original pretrained features alone are not sufficient for this task, and that the two proposed components contribute substantial additional gains on top of them.

AuthorsOpenReview ↗

Response to Reviewer 33W4: W2/Q - Model-selection protocol and the validation set

Response to Reviewer 33W4: W2/Q - Model-selection protocol and the validation set

Thank you for pointing this out — the partition protocol was not described in the paper. We follow the data partition scheme of DetailSemNet (Shih et al., 2024): the train/test writer splits for CEDAR, BHSig-Bengali, BHSig-Hindi and ChiSig are 50/5, 50/50, 100/60 and 829/829, respectively. All experiments use writer-disjoint splits.

In addition, we hold out a validation set from each training split for hyperparameter search. For CEDAR, BHSig-Bengali and BHSig-Hindi we hold out five writers. For ChiSig each writer provides only five genuine and five forged signatures, so we hold out 30 writers to obtain a validation set of comparable size. These validation writers are disjoint from the test writers.

The feature layers, the SSL configuration and the distance weights were selected on this validation data, and the test sets were used only for the final evaluation reported in the paper. We will state this protocol explicitly in the revision.

Response to Reviewer 33W4: W3 - Statistical reliability

Thank you for pointing out the importance of assessing stability across repeated splits. We evaluated the five-writer training setting over five random seeds and report EER as mean ± sample standard deviation with 95% confidence intervals. Every run draws an independent random writer selection, uses a writer-disjoint split, and keeps the architecture and training procedure unchanged.

Table R1. Five-writer training over five random writer-disjoint splits.

DatasetACC (%) ↑EER (%) ↓95% CI of EER
BHSig-Bengali93.63 ± 0.736.46 ± 0.76[5.52, 7.40]
BHSig-Hindi90.29 ± 0.569.78 ± 0.56[9.09, 10.48]
CEDAR90.11 ± 0.4510.15 ± 0.52[9.51, 10.80]
ChiSig88.49 ± 1.6311.75 ± 1.74[9.59, 13.92]

Table R2. EER (%) of each individual split.

SplitBengaliHindiCEDARChiSig
15.719.2110.7010.00
27.2610.009.6213.90
36.2110.459.6110.92
47.2710.0610.3013.33
55.859.1810.5510.62

On the CEDAR benchmark split. We also repeated the CEDAR benchmark protocol (50 training / 5 test writers) over five random splits, giving EER of 0.20 ± 0.15% (0.2174, 0.2174, 0.4348, 0.0720, 0.0720), with a 95% confidence interval of [0.02, 0.39]. Its upper end remains below the 0.58% reported by DetailSemNet, the strongest published result on CEDAR.

These results will be included in the revised manuscript.

Reviewer_Qdg1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SigDINO, a writer-independent offline signature verification method based on DINOv3. The authors first adapt DINOv3 using synthetic signature images through self-supervised learning. Patch features from intermediate transformer layers are enhanced using cross-layer residual information. Foreground patch tokens are selected and compared through bidirectional local structural matching to compute a verification distance between two signatures.

The method is evaluated on four signature datasets, including cross-language and limited-writer settings. The authors report improved EER over existing methods and provide ablation studies for several components.

Critical strengths
  1. The use of DINOv3 for offline signature verification is interesting. The reported EER results are promising across four datasets.
  2. The evaluation includes cross-language transfer and limited-writer settings. The ablation studies also provide useful evidence for the effectiveness of domain adaptation and local structural matching.
Critical weaknesses
  1. The methodological novelty appears limited. The self-supervised teacher–student adaptation follows established DINO methodology. The cross-layer enhancement is a simple linear combination of features. Local structural matching is also an established direction in signature verification. The paper does not sufficiently establish what fundamentally new learning principle is introduced.
  2. The contribution of the proposed components is not fully isolated. Table 6 indicates that much of the performance gain comes from the pretrained DINOv3 backbone. While the authors provide several ablations, the justification for the specific layer combination and residual formulation remains limited. A more controlled comparison with existing local matching approaches using the same backbone would help establish the contribution.
Questions for authors
  1. The presentation could be improved by explicitly explaining the tensor shapes and transformations. The teacher receives global crops of 224×224, while the student additionally receives local crops of 96×96. How are the different numbers of patch tokens handled? The authors should clarify the patch-token dimensions and how the teacher and student generate compatible output distributions.

  2. The authors use features from L17 and L23. Table 6 compares individual feature-extraction layers, but does not systematically examine different layer pairs. Can the authors provide an ablation comparing alternative pairs such as L14–L20, L14–L23, and L17–L20?

  3. Equation (5) defines the enhanced feature as (h_i=2\tilde p_i^{17}-\tilde p_i^{23}). Why is this particular formulation selected? Table 6 shows improvement with cross-layer enhancement, but alternative combinations and weighting coefficients should be examined.

  4. The authors use a grayscale ink map to select stroke-responsive patch tokens. However, its construction is not sufficiently explained. How are the pixel-level ink values aggregated into patch-level scores? What criterion determines foreground token selection? How are the normalized foreground weights in Equation (6) computed?

  5. The figure uses (n) aligned token pairs, whereas the formulation allows different numbers of selected tokens (n_1) and (n_2). The authors can clarify the soft correspondence process, bidirectional matching, and intermediate tensor dimensions.

  6. The authors report very low EER values, particularly 0.07% on CEDAR. However, threshold-dependent FAR and FRR behaviour is not shown. Can the authors provide ROC or DET curves and clarify the threshold-selection protocol? This is particularly relevant because CEDAR contains only five test writers.

  7. The authors should evaluate the model over multiple writer-disjoint train/test splits, keeping the architecture and training procedure unchanged (including L17, L23, Eq. 5 and other steps), and report mean performance with standard deviations. This is particularly important for CEDAR, where only five writers are used for testing, and for the five-writer training experiments.

10 comments
AuthorsOpenReview ↗

Response to Reviewer Qdg1: W1 - Novelty and isolation of component contributions

We sincerely thank you for these thoughtful comments on novelty and on the contribution of each component.

Few prior works have explored large-scale self-supervised models for offline signature verification. This requires bridging the distribution gap between natural images and signature images, where discriminative cues lie in local stroke structures rather than object-level semantics. Meanwhile, we would like to clarify why pre-trained representations alone cannot tackle the two challenges we focus on: limited training writers in downstream tasks and cross-lingual generalization, and we provide new experimental evidence for each of our two core designs.

1. Signature-domain adaptation for limited-writer training.

Signature-domain adaptation uses unlabeled synthetic signatures prior to downstream supervised training, with the aim of reducing reliance on a large set of labeled training writers.

We realize that one point in our ablation tables was not stated clearly. "Raw DINOv3" in Tables 6 and 7 retains our cross-layer feature processing and local structural matching module; only the signature-domain self-supervised adaptation is removed. Its numbers therefore reflect unadapted DINOv3 combined with our local structural matching, not raw DINOv3 global features. We will relabel this row in the revision.

Table R1. Effect of signature-domain adaptation under five training writers. Both settings use the same verification architecture. Reproduced from Table 7; EER (%), lower is better.

Backbone settingBengaliHindiCEDARChiSigAverage
Without signature-domain adaptation20.9211.5117.5816.3816.60
With signature-domain adaptation5.859.1810.5510.629.05

With the verification architecture held fixed, signature-domain adaptation lowers EER on all four datasets, reducing the average from 16.60% to 9.05% (-7.55 points). This provides direct evidence of its benefit under limited-writer supervision.

2. Local structural matching for cross-language transfer.

Local structural matching addresses the need to compare local stroke structures across writing systems. Aggregating each signature independently into a single vector before comparison can weaken subtle local stroke differences. SigDINO therefore preserves foreground patch representations, establishes bidirectional soft correspondences, and then measures local discrepancies and aggregates the matched relations, so that the resulting representation reflects interactions between the two signatures rather than each one in isolation.

To examine this design directly, we ran a head-replacement experiment. We kept the signature-adapted backbone, the cross-layer enhancement, and the foreground feature extractor unchanged, and replaced only the local structural matching module with an MLP verification head of comparable parameter count; verification then uses the cosine distance between independently computed signature embeddings.

Table R2. Verification-head comparison; EER (%), lower is better. Rows give the training dataset and head; columns give the test dataset. The "Local structural matching" rows are from Tables 1, 2 and 4.

Training setHeadBengaliHindiCEDAR
BengaliMLP2.1722.3910.58
BengaliLocal structural matching (ours)1.0311.755.87
HindiMLP10.384.9214.57
HindiLocal structural matching (ours)3.261.937.75
CEDARMLP13.3619.381.23
CEDARLocal structural matching (ours)7.7510.160.07

The MLP variant reaches an average in-domain EER of 2.77% over the three datasets and 15.11% over the six cross-language transfer directions; the corresponding figures for local structural matching are 1.01% and 7.76%, with lower EER in every one of the six transfer directions. The gap is modest in domain and much larger across languages, which is consistent with our motivation: preserving and comparing local correspondences before final aggregation matters most when the two signatures come from different writing systems.

3. Positioning of the contribution.

We would summarize our contribution as combining signature-domain representation adaptation with foreground local structural matching to address limitations that general-purpose pretrained representations exhibit on signature images: unlabeled signature adaptation supports training with few writers, while bidirectional local correspondences and matched-relation aggregation support cross-language verification. The ablations above indicate that the original pretrained features alone are not sufficient for this task, and that the two proposed components contribute substantial additional gains on top of them.

AuthorsOpenReview ↗

Response to Reviewer Qdg1: Q2/Q3 - Layer selection and the residual formulation in Equation (5)

Response to Reviewer Qdg1: Q2/Q3 - Layer selection and the residual formulation in Equation (5)

Thank you for this valuable question. Equation (5) is algebraically equivalent to 2p~i17−p~i232\tilde{p}_i^{17}-\tilde{p}_i^{23}. We adopt this residual formulation to clarify our design rationale: we build upon the representation at layer L17, and inject the difference between L17 and the final layer L23 as a corrective signal.

Why intermediate layers? Offline signature verification relies primarily on stroke geometry — spatial positions, connectivity and local layout — rather than on object-level semantic concepts. Analyses of pretrained visual backbone representations indicate that shallow layers retain texture detail, intermediate layers capture structural layout, and deep layers move toward abstract semantics. Intermediate layers therefore offer a favourable balance between spatial awareness and contextual structure, which benefits position-sensitive tasks (Amir et al., 2021). DINOv3 shows the same tendency: in its per-layer analysis, geometry-focused tasks perform better below the final layer, and its dense prediction pipeline aggregates several intermediate-layer features. These findings motivate the use of intermediate-layer features, while the exact indices are determined by ablation.

Why L17 and L23? Our ViT-L/16 backbone contains 24 Transformer blocks with zero-based indexing. L17 lies in the upper-middle range and L23 corresponds to the final block. Table R1 reports the ablation over alternative layer pairs and fusion strategies.

Table R1. Ablation on layer combinations and fusion strategies (EER, %). ChiSig is evaluated under the 1,658-writer protocol.

Feature configurationBengaliHindiCEDARChiSigAvg
Single layer
L171.652.380.141.671.46
L232.933.470.224.622.81
Simple fusion of L17 and L23
L17 + L232.042.080.143.681.99
L17 − L231.372.310.221.981.47
concat(L17, L23)2.312.220.431.711.67
Residual with other layer pairs
L14 + (L14 − L17)1.081.960.071.881.25
L14 + (L14 − L20)1.621.960.221.961.44
L14 + (L14 − L23)1.121.940.142.061.31
L17 + (L17 − L20)1.752.050.221.791.45
L20 + (L20 − L23)1.432.330.293.041.77
Ours
L17 + (L17 − L23)1.031.930.071.641.17

Our configuration achieves the lowest average EER on all four datasets. Using L17 alone gives 1.46%, and simple fusions of L17 and L23 — addition, subtraction, concatenation — all remain at or above that level, so the gain does not come from merely having access to L23. Adopting the residual form lowers the average across all layer pairs tested, and among them L17 with a correction from L23 is the strongest: holding the base at L17, taking the correction from L23 rather than L20 reduces the average from 1.45% to 1.17%. This supports the design: L17 is retained as the primary local representation, and a cross-layer correction derived from a deeper block enhances it.

Revised Figure 3. To further verify and visually demonstrate our design, we have updated Figure 3 in the main text to an occlusion-sensitivity map. We mask each input patch sequentially and record the variation of matching distance; locations with larger distance shifts are highlighted with higher heat response. This visualization intuitively illustrates how we retain L17 as the core local representation and leverage L23 to produce a cross-layer correction term for representation enhancement.

AuthorsOpenReview ↗

Response to Reviewer Qdg1: Q1 - Tensor shapes and compatibility between global and local views

Thank you for this suggestion — we will add the full set of shapes to Sec. 3.2. We describe the adaptation stage below.

For DINOv3 ViT-L/16, the patch size is 16×1616\times16 and the token feature dimension is 1024. For a batch of size BB, a 224×224224\times224 global view produces patch tokens of shape B×196×1024B\times196\times1024, and a 96×9696\times96 local view produces patch tokens of shape B×36×1024B\times36\times1024, excluding prefix tokens. Each view is processed separately, and from each we take the CLS representation of shape B×1024B\times1024, which has the same shape for every view regardless of crop size.

The teacher and student DINO heads map their respective CLS representations to logits of shape B×KB\times K, where KK is their common output dimension. Teacher targets are obtained through centering, temperature scaling, and softmax; student log-probabilities are computed through temperature scaling and log-softmax. The cross-entropy in Eq. (1) is thus computed over the same KK-dimensional output space for every view.

The teacher processes two global views, while the student processes both global views and four local views. After separate forward passes, their logits are concatenated along the batch dimension into 2B×K2B\times K and 6B×K6B\times K tensors, respectively, and then split by view for loss computation. Each teacher prediction supervises the five student views other than the identical view, yielding ten cross-view loss terms, averaged over the view pairs and batch.

Patch tokens enter the pipeline in the verification stage, where every input is resized to 448×448448\times448, giving a uniform $28\times28 $ token grid. The L17 and L23 tokens share this identical spatial layout, which is what makes the element-wise cross-layer residual in Eq. (5) well defined.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Thank you for the constructive feedback. We summarize the main changes in the revised manuscript below.

  • Method and ablation clarifications. We clarified how teacher and student predictions from global and local crops are compared in a shared output space (Section 3.2), specified that foreground weights are distributed uniformly over the selected tokens within each signature (Section 3.4), and updated the visualization of cross-layer residual enhancement (Figure 3). We also revised the ablation study to compare the same DINOv3 backbone with and without signature-domain self-supervised adaptation (Section 4.5, Tables 6 and 7).
  • ChiSig evaluation. We updated the main benchmark and corresponding ablation experiments with corrected ChiSig results under the full 1,658-writer protocol (Tables 3 and 6). Appendix A.4 (Table 9) further reports bidirectional cross-language transfer between ChiSig and the other three datasets under both the 395-writer balanced protocol and the full 1,658-writer protocol. The two protocols provide complementary views of model performance: the 395-writer protocol measures cross-language performance under class balance, while the 1,658-writer protocol reflects the practical setting in which forged signatures are scarce. We also clarified that only a subset of ChiSig writers provide both genuine signatures and skilled forgeries (Section 4.1) and explained how the five training writers are selected for ChiSig (Appendix A.2).
  • Additional appendix experiments. We added a comparison between local structural matching and an MLP verification head for cross-language transfer (A.5, Table 10), five-writer training results across five random writer-disjoint splits (A.6, Table 11), ROC and DET curves for all four datasets (A.7, Figure 7), and a sensitivity analysis of the cross-layer residual coefficient (A.12, Table 18). Substantive revisions are marked in red in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.