acceptodds
Under review as a conference paper at ICLR 2027

Ridges Beyond the Window: Structure-First Completion of Contactless Fingerprints from the Capture Sweep

Abstract

A contactless fingerprint captured in the wild is both degraded and partial, because a phone camera sees a blurred, noisy patch of the fingertip rather than a clean rolled impression. We present SweepFill, a single network that restores what such a capture shows and generates ridges of the correct flow where it shows nothing. Each capture is a short video of the turning finger, which we call a sweep, so a part hidden in one frame often appears in another, and no published completion method uses this. Because the exact position of ridges (their phase) beyond the visible window is not determined by the observation, while their direction (flow) and spacing (period) are, SweepFill predicts structure before pixels. A head on the bottleneck predicts ridge flow and frequency for the whole print, a phase-free spectral loss lets the decoder draw ridges without being punished for phase it cannot know, and joint attention over other frames of the same sweep turns much of the generation into registration, all learned by regression alone. We also show that inserting the mask token after the positional encoding makes every hidden token of a masked-bottleneck transformer identical, a failure not specific to fingerprints. We evaluate with two fingerprint matchers under a cross-session protocol. When 30% of the print is visible, SweepFill reduces hidden-region orientation error from 20.07 to 11.01 degrees and raises ridge contrast from 0.15 to 0.67 relative to a DiT-UNet restorer trained on more samples, and it lowers the deep matcher’s EER from 26.0% to 11.6% on held-out MMFV subjects. Completing a poor frame before fusing scores over the other frames raises fused rank-1 by +8.3 and +11.5 points when 50% and 70% of the print is visible (not significantly at 30%), and an in-domain fine-tune improves both matchers on a second device. The generated flow is within 10.5 degrees RMSE of the clean frame, but its phase is uncorrelated with it, and only 21% of true minutiae are recovered, so the completions are plausible rather than reconstructed.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
contactless fingerprints, partial-print completion, ridge orientation field, biometric evaluation

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_5QtoOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary
  • The paper deals with training a completion model to recover a fingerprint with partial visibility from phone cameras.
  • Their main contribution is to utilize other frames to complete details that are missing in the base (degraded) frame.
  • Their method contains three components: (1) predicting the ridge structure first (before pixel reconstruction), (2) phase-free spectral loss which is based on comparing the frequency content instead of the pixels, and (3) attention of additional reference images from the sweep. 
The experimental section uses two datasets: MMFV and ISPFDv2, and two print matchers.
Critical strengths
  • The idea of incorporating additional frames to complete missing details from the degraded frame sounds interesting and promising
  • The results at Table 1 shows clear trend of improvement when using SweepFill with varied ‘w’
Critical weaknesses
  • I find the paper hard to follow. The authors could simplify phrases and better organizing the sections. Instead, long and complex sentences made the paragraphs hard to understand. See for example, lines 279 - 282: “Because everyMMFV frame shows the whole fingertip, the window is synthetic and the references always contain the hidden region, the favorable end of the sweep setting, and we bound the other end by degrading the references (Table 13, Appendix E), removing them (Table 35, Appendix K.3) and occluding them(Appendix D.1). “. I would clarify the paragraphs and make the paper more readable. Currently the paper is not clear or holistic enough for publication.
  • It is not clear why diffusion models based inpainting cannot be the basis for fingerprint reconstruction? One can input the full video sweep to the model and it could be a valid baseline. The authors did explained in Appendix I why there is no diffusion baseline. However, applying diffusion models conditions with additional frames from the sweep can be the right baseline to compared to.
  • Important ablations appear very late in the paper’s appendix. For example, a comparison to multi-frames fusion directly. In Table 1 the comparison is only to degrade-probe. I would expect to see such a comparison in the main table.
Questions for authors
  • No questions
5 comments
AuthorsOpenReview ↗

Authors Response (1/3)

We thank the reviewer for the concrete suggestions on readability and on the order of the results. Following the reviewer's suggestions, we ran each requested analysis, and all of them are now in the paper. For clarity, the main text now uses shorter sentences and moves the key comparisons into Tables 1 and 2. Below, EER is the equal error rate (lower is better), rank-1 is the share of probes whose top match is the right person, and percentages of the print visible describe the probe.

  • Weakness 1: the main text now uses shorter sentences and moves the key comparisons into Tables 1 and 2.
  • Weakness 2: SweepFill reaches 79.2% rank-1 with 30% of the print visible, against 54.2% for DiffuEraser and 29.2% for LeftRefill.
  • Weakness 3: Table 1 now includes multi-frame fusion, and fusing with the output of SweepFill gives the lowest EER among the degraded-frame options (8.3%).

Weakness 1

I find the paper hard to follow. …

We shortened long sentences throughout the main text and moved the key comparisons into the main tables. The main text now uses shorter sentences, and Tables 1 and 2 carry the comparisons a reader needs. The sentence the reviewer quoted now reads "Every MMFV frame shows the whole fingertip, so the window is synthetic and the references always contain the hidden region. This is the favorable case for sweep references. We test less favorable cases by degrading the references (Table 15, Appendix E), removing them (Table 38, Appendix K.3) and occluding them (Appendix D.1)." Table 1 now carries the comparison with multi-frame fusion, and the new Table 2 in Section 5.2 puts LaMa, TransRef, LeftRefill and DiffuEraser on the same inputs as SweepFill, using the extra page allowed during the discussion.

AuthorsOpenReview ↗

Authors Response (2/3)

Weakness 2

It is not clear why diffusion models based inpainting cannot be the basis for fingerprint reconstruction? …

We trained the prompts of LeftRefill (Cao et al., CVPR 2024), a reference-conditioned diffusion inpainter on frozen Stable Diffusion 2, on our crops with the sweep frames as its references, and ran the video diffusion inpainter DiffuEraser (Li et al., 2025) with 21 sweep frames as input. With 30% of the print visible, SweepFill reaches 79.2% rank-1, against 54.2% for DiffuEraser and 29.2% for LeftRefill, and it also leads both on equal error rate, orientation error and unsupported minutiae. Its rank-1 lead over DiffuEraser is significant over subjects. DiffuEraser with its released weights is not significantly different from SweepFill trained at the equal budget.

All rows

SweepFill at the larger budget leads in every column except missed minutiae. Missed is the share of true hidden minutiae with no generated counterpart, and unsupported the share of generated minutiae there with no true counterpart.

30% visibleEER% ↓rank-1% ↑minutiae EER% ↓orientation error° ↓missed / unsupported% ↓
LeftRefill (equal budget), guidance 2.537.59.442.827.177.2 / 86.1
LeftRefill (equal budget), guidance 1.028.129.239.725.273.3 / 82.2
LeftRefill (1.9× budget), guidance 1.031.221.938.625.677.0 / 82.6
DiffuEraser, released weights, 21 frames17.754.229.015.776.7 / 66.0
SweepFill (equal budget), 2 frames18.857.330.414.193.6 / 75.0
SweepFill (larger budget), 2 frames11.679.223.210.978.7 / 54.2

DiffuEraser

The published SweepFill raises rank-1 by 25.0 points over DiffuEraser ([+10.4, +38.5], p=0.031) and the true accept rate at a 1% false-accept rate by 18.8 points ([+5.2, +29.2]), with a lower EER (11.6%, jackknife of the difference [-9.3, -2.8]) and orientation error (10.9 against 15.7°, point estimates). At the equal budget DiffuEraser is not significantly different from SweepFill (17.7 against 18.8% EER, jackknife [-5.3, +7.4], rank-1 54.2 against 57.3%, p=0.81). Each probe sits in a 22-frame clip with 21 frames of the other two sweeps of the same session, including the two frames SweepFill uses, only the probe frame is masked, and DiffuEraser leaves the visible window as it is. Of its hidden-region minutiae 66.0% are unsupported, between the two SweepFill models (75.0 and 54.2%).

LeftRefill

LeftRefill does not lower EER (28.1 against 26.0% for the raw probe, 31.2% after 1.9× the budget), its rank-1 is above the raw probe's 21.9% in all four sampling seeds (25.0 to 33.3%, not tested over subjects), and its orientation error is 25.2° against 14.1° for SweepFill at the equal budget. Its unsupported-minutiae rate (82.2%) is above that of SweepFill even with another subject's frames (60.7%). With the authors' guidance of 2.5 it oversaturates its outputs, and guidance 1.0 also scores best on 8 training subjects.

LeftRefill with more visible

Trained for 1.9× the budget, LeftRefill does not lower the raw probe's EER with 50 or 70% visible, and its rank-1 is equal at 50% and higher at 70% (63.5 against 59.4%) (point estimates, no paired test).

deep matcherEER% ↓rank-1% ↑
raw probe, 50% visible20.853.1
LeftRefill 1.9×, 50% visible20.853.1
raw probe, 70% visible15.859.4
LeftRefill 1×, 70% visible14.666.7
LeftRefill 1.9×, 70% visible15.863.5

Section 5.2 and Appendix I now report both.

AuthorsOpenReview ↗

Authors Response (3/3)

Weakness 3

Important ablations appear very late in the paper’s appendix. …

We added this comparison to Table 1 and fused the other frames, degraded like the probe, with the raw probe and with the output of SweepFill. Fusing them with the output of SweepFill gives the lowest EER among the degraded-frame options, 8.3% with 30% of the print visible, against 12.5% for the frames alone and 10.4% for the degraded probe fused with them. Against the frames alone the change holds over subjects, and against the fused raw probe it does not.

Fusion and the other baselines

Fusing the degraded frames with the output of SweepFill gives the lowest EER, and only matching the two clean frames directly is better (Table 35). Deep matcher, 30% of the print visible.

probeEER% ↓rank-1% ↑
degraded probe26.021.9
SweepFill, references switched off at test17.655.2
mosaicking from the two sweep frames16.754.2
two degraded frames alone12.578.1
degraded probe fused with them10.478.1
SweepFill from the degraded frames14.477.1
SweepFill from the degraded frames, fused with them8.380.2
SweepFill with clean references11.679.2
two clean frames matched directly6.193.8

Table 1 now lists these rows with the other baselines, and the new Table 2 in Section 5.2 holds the comparison with LaMa and the reference-guided inpainters.

SweepFill leads the diffusion inpainters given the sweep, and Table 1 and the new Table 2 now carry the comparisons the reviewer asked to see in the main text.

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

SweepFill restores and completes partial contactless fingerprints using structure supervision, a phase-free spectral loss, and attention over two reference frames. Cross-session experiments show improved ridge structure and recognition. Minutiae and phase analyses distinguish plausible completion from recovery of true fingerprint detail.

Critical strengths
  1. The spectral objective is well motivated and empirically effective. Table 29 separates its contribution from the FFC architecture and shows substantially better hidden-region ridge orientation and contrast. This provides a useful approach to deterministic completion when exact phase recovery is uncertain.
  2. The evaluation distinguishes visual plausibility from biometric fidelity. Cross-session matching, region-composition controls, and minutiae analysis clarify both the usefulness and limitations of generated ridges. These findings are informative beyond image-quality improvements.
Critical weaknesses

The incremental value over using the available reference frames remains insufficiently established. At 30% visibility, Table 35 improves fused EER from 10.42% to 8.33%, but the paired difference interval includes zero. The positive fused rank-1 results at 50% and 70% visibility are encouraging, although their uncertainty is assessed over fingers rather than subjects. An established learned reference-guided baseline, such as TransRef, is also missing. These gaps limit attribution of the gains to the proposed design.

Questions for authors

Do the positive fusion results in Table 35 remain stable under subject-level uncertainty estimates? Is there a matched-input comparison with an established reference-guided inpainter that clarifies the proposed method’s incremental benefit?

4 comments
AuthorsOpenReview ↗

Authors Response (1/3)

We thank the reviewer for testing the method against the reference frames themselves. Following the reviewer's suggestions, we ran each requested analysis, and all of them are now in the paper. Below, EER is the equal error rate (lower is better), rank-1 is the share of probes whose top match is the right person, and percentages of the print visible describe the probe.

  • Weakness 1: adding the completed probe lowers the two frames' EER from 12.5 to 8.3% with 30% of the print visible, significant over subjects.
  • Question 1: the fused rank-1 gain with 70% visible holds over subjects. Against TransRef, SweepFill has the lower orientation error and a higher minutiae-matcher TAR, while TransRef reaches a lower EER.
  • Question 1 (training seeds): on two further training seeds of both models, SweepFill keeps the higher rank-1 on every seed, and TransRef keeps the lower EER.

Weakness 1

The incremental value over using the available reference frames remains insufficiently established. …

We fused the completed probe with the two other frames, degraded like the probe, and compared it with the frames alone and with the raw probe fused with them. Adding the completed probe lowers the frames' equal error rate from 12.5 to 8.3% with 30% of the print visible, significant over subjects, and with 70% visible it also raises rank-1 by 14.6 points. Against fusing the raw probe instead, the further gain with 30% visible is not significantly different from zero on the published split, while on two new subject splits completing before fusion lowers the fused EER by 4.8 and 3.3 points (by point estimate, significant over subjects on the second split only). TransRef is in our reply to Question 1.

Fusion with the two frames

Adding the completed probe gives the lowest fused EER with 30% of the print visible, while completion without fusion is no better than the frames alone. Deep matcher, frames degraded like the probe.

30% visibleEER% ↓rank-1% ↑
two frames alone12.578.1
raw probe fused with them10.478.1
completed probe fused with them8.380.2
completion without fusion14.477.1

The EER change from adding the completed probe has leave-one-subject-out jackknife interval [-7.9, -0.4] (p=0.034), and rank-1 moves from 78.1 to 80.2% (p=0.63). With 70% visible the change is -7.3 EER points (jackknife [-13.2, -1.3]) and +14.6 rank-1 points (p=0.031). For a single frame with no video, completion lowers EER from 26.0 to 17.6% and raises rank-1 from 21.9 to 55.2% (both p ≤ 0.012 over subjects), and the single-frame gains remain significant over subjects at every visibility (p ≤ 0.016).

Section 5.3 and Table 1 now report these comparisons.

Question 1

Do the positive fusion results in Table 35 remain stable under subject-level uncertainty estimates?

We added subject-level tests on the published split and on two new subject splits, and we trained TransRef on our data with the same inputs as SweepFill.

Subject-level tests

The fused rank-1 gain with 70% of the print visible holds over subjects (+11.5 points, p=0.031), and on the two new subject splits completing before fusion lowers the fused EER at every visibility by point estimate and never lowers fused rank-1. With 50 and 30% visible the fused rank-1 gains on the published split are not significantly different from zero.

Subject-level tests per split

On the published split the fused rank-1 gains with 50 and 30% visible are +8.3 (p=0.062) and +2.1 (p=0.63), and no fused EER change from completing has a jackknife interval excluding zero. On the two new splits the fused EER change with 30% visible is -4.8 (jackknife [-10.5, +1.0]) and -3.3 ([-6.3, -0.3]), and the fused rank-1 gain with 50% visible is +8.3 on the second (p=0.031) and +9.5 on the first (p=0.125).

Section 5.3 now reports these tests.

AuthorsOpenReview ↗

Authors (2/3)

Question 1

Is there a matched-input comparison with an established reference-guided inpainter that clarifies the proposed method’s incremental benefit?

We trained TransRef (Liu et al., Neurocomputing 2025) with its own losses on our crops, windows and sweep frames, from its released weights at the equal budget and from scratch at both budgets, and gave it the first reference frame of SweepFill. TransRef reaches a lower EER, while SweepFill has the lower hidden-region orientation error (10.9 against 12.6°, significant over subjects) and, on the 14 subjects of the two new splits never used for selection with 30% visible, a minutiae-matcher TAR (accepted genuine matches at a 1% false-accept rate) 15.5 points higher (significant after Bonferroni correction over 18 tests although not planned in advance).

Published split

TransRef leads on EER, and SweepFill leads on rank-1 with 30% visible.

published split, 30 / 50 / 70% visibleEER% ↓TAR% ↑rank-1% ↑
SweepFill (larger budget)11.6 / 9.4 / 7.357.3 / 64.6 / 66.779.2 / 83.3 / 88.5
TransRef from scratch (larger budget)8.6 / 5.1 / 4.460.4 / 81.2 / 86.572.9 / 90.6 / 93.8
TransRef pretrained (equal budget)8.3 / 5.2 / 6.255.2 / 83.3 / 82.369.8 / 85.4 / 88.5

EER against TransRef from scratch

TransRef from scratch has the lower EER at every visibility. On the published split the difference is significant over identities with 50 and 70% visible and over subjects with 70% only, and with 30% visible it is not significantly different (+3.1, jackknife [-1.5, +7.7], p=0.16). On the two new splits, over their 14 subjects never used for selection, the subject bootstrap intervals exclude zero at every visibility. Differences are SweepFill minus TransRef, so a positive EER difference favors TransRef.

EER difference ↓30% visible50% visible70% visible
published split+3.1+4.2, [+2.0, +8.4]+2.9, [+0.6, +5.2], jackknife [+0.1, +5.7]
two new splits+7.1, [+3.9, +10.0]+4.8, [+1.9, +7.1]+5.3, [+2.5, +7.7]

Rank-1 and TAR against TransRef from scratch

Under the deep matcher no rank-1 difference is significant on any split (p ≥ 0.25) or on the two new splits together (p ≥ 0.13). On the 14 subjects of the two new splits never used for selection, with 30% visible, SweepFill's minutiae-matcher TAR is 15.5 points higher (subject bootstrap [+6.0, +20.2], exact permutation p=0.0002, not planned in advance). It is +9.4 (p=0.13) on the published split.

rank-1 difference ↑30% visible50% visible70% visible
published split+6.2-7.3-5.2
new split 1+8.3-6.0-9.5
new split 2+4.2+2.5-5.8

Orientation error and sweep frames

SweepFill's hidden-region orientation error is 10.9°, against 12.6° for TransRef from scratch (p=0.016) and 13.7° pretrained (p=0.008, the smallest value attainable with 8 subjects). Trained without its sweep frames at the larger budget, SweepFill's rank-1 with 30% visible is 22.9 points lower (p=0.031), and with another subject's frames it falls to 25.1% EER and 41.7% rank-1 while its orientation error rises from 10.9 to 15.7° (p=0.008). TransRef from scratch keeps its identity scores with another subject's frame on the published split (EER 8.3 against 8.6%, jackknife [-3.4, +2.9], rank-1 p=1.0) and on the first new split (EER change 0.0 at every visibility), except on the second new split with 30% visible, where its EER rises by 4.2 points (p=0.029, and +2.5, p=0.077, on the subjects never used for selection).

Hidden minutiae

Every method recovers more hidden minutiae than an other-identity null, another identity's completion by the same method (SweepFill 21.3 against 9.6%, TransRef 48.4 against 28.1%, p=0.008 for both), the null scales with how many minutiae a method places, so TransRef's higher recovered rate is largely density, and SweepFill places fewer unsupported minutiae (54.2 against 65.9%).

Pretraining on contact prints

Pretrained on NIST SD302 contact prints (4.3M sample visits) and fine-tuned at the equal budget, SweepFill reaches 78.1% rank-1 (+20.8 over training from scratch, p=0.031), against 69.8% for the pretrained TransRef (p=0.20), with EER 13.5 against 8.3%, not significantly different (jackknife [-3.0, +13.4]). Section 5.2 reports TransRef at all visibilities and on both new splits and the reference controls, Section 5.4 the minutiae null, and Appendix E.5 the SD302 pretraining.

The abstract and limitations now state the comparison with TransRef, and Section 5.2 reports it.

AuthorsOpenReview ↗

Authors Response (3/3)

Question 1 (training seeds)

Is there a matched-input comparison with an established reference-guided inpainter that clarifies the proposed method’s incremental benefit?

We trained two further seeds of SweepFill and of TransRef from scratch, both at the larger training budget and with the same settings as the published runs. The advantages of SweepFill hold on every seed. With 30% of the print visible on the published split, every SweepFill seed reaches a higher deep-matcher rank-1 than every TransRef seed, and SweepFill keeps the higher TAR and rank-1 under the minutiae matcher. TransRef keeps the lower deep-matcher EER, as in our reply to Question 1.

published split, 30% visible, mean ± SD over 3 seedsSweepFillTransRef
deep matcher, rank-1% ↑79.5 ± 0.670.8 ± 2.8
minutiae matcher, TAR at 1% false-accept rate, % ↑62.2 ± 1.648.3 ± 4.8
minutiae matcher, rank-1% ↑67.7 ± 2.855.9 ± 3.7
deep matcher, EER% ↓12.3 ± 1.09.5 ± 0.9

Completing the probe adds to the reference frames, most clearly with more of the print visible and on the single frame, and SweepFill's sweep references set a hidden-region orientation error that TransRef does not reach.

AuthorsOpenReview ↗

With our thanks, and a gentle follow-up

We thank the reviewer again for the review. In brief, our reply shows that adding the completed probe lowers the two frames' EER from 12.5 to 8.3% with 30% of the print visible, significant over subjects, and that against TransRef trained on the same inputs, SweepFill keeps the lower hidden-region orientation error and a higher rank-1 on every one of three training seeds. Since the discussion period ends soon, we would be grateful to hear whether our reply addresses the reviewer's concerns, and we would be glad to clarify any remaining point.

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

The paper presents SweepFill for restoring degraded contactless fingerprints and completing masked regions using other frames from the same capture sweep. It combines ridge-structure prediction, a local amplitude-spectrum loss that does not require matching ridge phase, and joint attention to reference frames. The evaluation on MMFV and ISPFDv2 considers image quality, cross-session identity matching, minutiae fidelity, and the effects of incorrect references. The results show useful recognition gains under the evaluated conditions, while the minutiae and phase analyses appropriately distinguish plausible ridge completion from recovery of the true missing fingerprint detail.

Critical strengths
  1. Well-motivated use of structure and capture-sweep information. Ridge prediction and the amplitude-spectrum loss suit the completion task. Ablations and comparisons using the same reference frames help assess the benefit of capture-sweep information beyond single-image restoration.
  2. Evaluation considers recognition, not just appearance. Cross-session matching and minutiae, phase, and wrong-reference analyses test whether completion improves recognition and what information it recovers. The paper also clearly acknowledges that generated ridges may differ from the true missing details.
Critical weaknesses
  1. Independent evaluation of the full model is unclear. The MMFV subjects used for the main results also guided model and parameter selection. Appendix B.2 provides independent evaluations, but only for models trained with less data and fewer epochs. Please evaluate the full model on subjects not used for model or parameter selection.
  2. The comparison with LaMa does not fully isolate the source of improvement. SweepFill uses more training data, more epochs, and additional reference frames. Existing ablations provide useful evidence, but these differences complicate attribution of its advantage over LaMa. Please compare equal-budget and reference-free variants with LaMa on the same test inputs.
Questions for authors
  1. LaMa and SweepFill differ in training data volume, training epochs, and reference-frame inputs, so their performance gap may not arise solely from algorithm design. Please provide comparisons under similar training conditions or report LaMa’s performance after longer training, and use reference-frame ablations to clarify the contribution of each factor.

  2. The experiments mainly simulate partial inputs by artificially masking complete fingerprints. Further evaluation on genuinely partial fingerprint captures would strengthen the validation. In addition, ISPFDv2 preprocessing excludes approximately half of the candidate images. Please report the proportion excluded for each reason and discuss how this filtering affects practical applicability.

  3. Appendix A indicates that the subjects evaluated in Table 1 were also used for model selection and parameter tuning. Although Appendix B.2 evaluates two additional subject groups, those experiments use models trained with smaller budgets. Please clarify whether the final full model was evaluated on a separate, independent test set.

  4. Section 3.5 and Table 3 describe different patch-selection criteria for the spectral loss. Please clarify the region over which each proportion is calculated, whether both criteria are applied jointly, and consistently state the rule actually used during training

7 comments
AuthorsOpenReview ↗

Authors Response (1/4)

We thank the reviewer for the specific questions on training budgets and on the independence of the evaluation. Following the reviewer's suggestions, we ran each requested analysis, and all of them are now in the rebuttal and paper. For real partial captures, we also evaluated the ISPFDv2 images that fail the quality gate (our reply to Question 2). Below, EER is the equal error rate (lower is better), rank-1 is the share of probes whose top match is the right person, and percentages of the print visible describe the probe.

  • Weakness 1: on two new subject splits, completion lowers EER by 13.6 and 9.5 points and raises rank-1 by 47.2 and 49.0 points with 30% of the print visible.
  • Weakness 2 and Question 1: at the larger budget SweepFill leads LaMa by 31.2 rank-1 points, 22.9 of them from the reference frames, while longer training changes LaMa little.
  • Question 2: on 425 rejected ISPFDv2 images, completion still lowers EER from 35.6 to 28.0%.
  • Questions 3 and 4: the full model is now evaluated on independent subjects, and Table 3 states the single training rule.

Weakness 1

Independent evaluation of the full model is unclear. …

We retrained the full model with the same recipe on two new subject splits and scored it once, at its final epoch, on the 14 subjects never used for any choice. With 30% of the print visible, completion lowers the deep matcher's equal error rate by 13.6 and 9.5 points, and raises rank-1 by 47.2 and 49.0 points. All four gains are significant over subjects, and they are somewhat smaller than on the published split while holding on both new splits.

New splits 1 and 2 are splits B and C of Appendix B.2.

Tests over subjects

Every gain on the subjects never used for selection excludes zero over subjects. Differences are completion minus the raw probe, with 30% of the print visible.

deep matcherEER change ↓intervaljackkniferank-1 change ↑p over subjects
new split 1, 6 subjects-13.6[-21.7, -5.7][-21.8, -5.5]+47.20.031
new split 2, 8 subjects-9.5[-14.0, -4.4][-18.0, -0.9]+49.00.008

Results per split

SweepFill lowers EER and raises rank-1 on every split, with references switched off it still reaches 22.3 and 19.6% EER on the subjects of the two new splits never used for selection, and the gains are 1 to 5 EER points and 8 to 10 rank-1 points smaller than on the published split.

30% visible, deep matcheridentitiesEER% rawEER% SweepFillrank-1% rawrank-1% SweepFill
published split (Table 1)3226.011.621.979.2
new split 1, all subjects2833.319.011.967.9
new split 1, never used for selection2434.520.815.362.5
new split 2, all subjects4026.816.020.070.0
new split 2, never used for selection3227.217.719.868.8

Minutiae matcher

The minutiae matcher points the same way on both new splits, though on the second its EER gain does not hold at the subject level (Appendix B.2).

minutiae matcher, 30% visibleEER% raw to SweepFill ↓intervaljackkniferank-1 gain ↑p
new split 1, never used for selection40.4 to 27.2[-19.9, -3.2][-24.8, -1.6]+37.50.031
new split 2, never used for selection35.8 to 23.1[-18.8, -5.1][-29.2, +3.8]+24.00.016

Earlier evidence

At the equal budget, completion already lowers EER and raises rank-1 on all three splits (Table 6), and on ISPFDv2 subjects outside every choice the published model lowers EER from 28.8 to 17.9% (Table 37).

30% visible, equal budgetEER% raw to SweepFill ↓rank-1% raw to SweepFill ↑
published split29.1 to 20.027.1 to 58.3
new split 132.0 to 20.215.5 to 47.6
new split 225.8 to 22.522.5 to 36.7

Appendix B.2 now reports both matchers at every visibility and states how many subjects each split shares with the published split, and Section 5 now describes the Table 1 subjects as held out from training.

AuthorsOpenReview ↗

Author Response (2/4)

Weakness 2 and Question 1

The comparison with LaMa does not fully isolate the source of improvement. …

LaMa and SweepFill differ in training data volume, training epochs, and reference-frame inputs, so their performance gap may not arise …

We trained LaMa with its own losses and SweepFill without reference frames, each at the equal budget and at the larger training budget of SweepFill (8.1 times the training samples), on the 96 probes of Table 1. At the larger budget and with 30% of the print visible, SweepFill's rank-1 is 31.2 points above LaMa's, and its two reference frames account for 22.9 of those points, significant over subjects. Without reference frames SweepFill is not significantly different from LaMa with 30% visible and leads it with 70% visible, and at the equal budget the reference frames already raise rank-1 significantly. We also fine-tuned LaMa with its own losses for that budget and repeated the reference-frame ablation at both budgets.

All rows use the same 96 probes with 30% of the print visible, and the budget is the equal budget (1×) or the larger budget of SweepFill (8.1×). Lower orientation error and EER and higher rank-1 are better, band contrast is best at 1, and SweepFill at the larger budget leads on orientation error and rank-1.

method (budget)refs.orientation error° ↓band contrastEER% ↓rank-1% ↑minutiae EER% ↓
LaMa, our losses (1×)020.560.11718.640.632.8
LaMa, its own losses (1×)016.810.71917.946.933.6
LaMa, its own losses (8.1×)016.610.68115.647.927.5
SweepFill, no refs (1×)016.510.57220.044.826.3
SweepFill, 2 refs (1×)214.070.64918.857.330.4
SweepFill, no refs (8.1×)013.850.59515.656.222.4
SweepFill, references off at test (8.1×)014.070.58417.655.227.8
SweepFill (8.1×)210.900.71511.679.223.2

Orientation error is measured in the hidden region, and band contrast is the ridge-band contrast of the output over that of the ground truth.

Tests

The reference frames and the larger budget give significant rank-1 gains, and SweepFill at the larger budget is also significantly lower than LaMa in EER over subjects with 50 and 70% visible (jackknife [-5.9, -0.4] and [-5.5, -0.8], 3.1 points). Differences are SweepFill minus the comparison.

comparison (30% visible unless noted)resulttest
two references, larger budgetrank-1 +22.9, [+11.5, +35.4], EER -4.0, [-8.3, +2.9], jackknife [-12.5, +4.5]p=0.031
two references, larger budget, 50 and 70% visiblerank-1 +1.0 and +2.1, EER -4.1 and -2.1not significant
SweepFill against LaMa, larger budgetrank-1 +31.2, [+17.7, +44.8], EER -4.0, [-8.0, +1.9]p=0.039
no references against LaMa, larger budgetrank-1 +8.3p=0.38
no references against LaMa, larger budget, 70% visiblerank-1 +10.4p=0.031
two references, equal budgetrank-1 44.8 to 57.3%, orientation error 16.5 to 14.1°p=0.031
SweepFill against LaMa, equal budgetEER +0.9, [-7.3, +6.4], jackknife [-4.6, +6.4], rank-1 +10.4p=0.22
longer training, LaMaEER -2.3, [-7.6, +2.7], rank-1 +1.0, [-8.3, +10.4]not significant
longer training, SweepFillEER -7.1, [-9.8, -0.3], rank-1 +21.9p=0.031

The model trained without references is not significantly different at any visibility from the two-reference model with its references removed at test. Without references SweepFill has the lower minutiae EER than LaMa (26.3 against 33.6% at the equal budget and 22.4 against 27.5% at the larger budget, point estimates) and about the same orientation error at the equal budget (16.5 against 16.8°), which LaMa reaches only with its adversarial and perceptual losses (20.56° with ours). LaMa has 51.0M parameters against 126.2M for SweepFill (Appendices A and H).

Section 5.2 and Table 2 now report this.

AuthorsOpenReview ↗

Author Response (3/4)

Question 2

The experiments mainly simulate partial inputs by artificially masking complete fingerprints. …

We evaluated the method on real captures that fail the ISPFDv2 quality gate, broke the rejections down by reason and capture condition, and scored the rejected images that still give a crop. On these 425 rejected images, the fine-tuned model lowers the deep matcher's EER from 35.6 to 28.0% with 30% of the print visible. MMFV has no genuinely partial captures with a cross-session gallery (Appendix D.2), and the ISPFDv2 images rejected for a tip at the border or low fill are mostly segmentation and alignment failures (background, knuckles, or whole fingertips placed low in the crop), not prints cut by the frame. Section 6 now states that real partial captures with a cross-session gallery are the missing evaluation.

ISPFDv2 exclusions

The gate rejects 51.9% of the instances it examines, mostly because no finger is segmented (24.5%) or no ridge period can be measured (21.6%), and a plain indoor background passes about 84% of the time.

Rejection by reason

Of the 1,911 instances the gate examined, 991 (51.9%) were rejected, mostly because no finger was segmented.

rejected byshare of instances examined, %
no finger segmented24.5
no measurable ridge period21.6
fill below half4.0
fingertip at the border1.5
blur0.3

Rejection by capture condition

Rejection rises from 15.8% for white-background indoor captures to 75.0% for natural-background outdoor ones (Appendix B). Every method sees the same accepted crops (371 of the 460 accepted crops are indoor), so the filter favors no row.

background, illuminationrejected% of all images of the 96 attempted identities
white, indoor15.8
white, outdoor46.7
natural, indoor51.0
natural, outdoor75.0

Images outside the gate

On all 425 outside-gate images (66 identities) the fine-tuned model lowers EER from 35.6 to 28.0% ([-12.5, -2.3]). On the 323 that failed first on the ridge test, real fingertips with faint ridges (233 images of 45 identities without the seven identities below), it keeps about half of its gain on accepted images with 30% visible and most of it with 50% visible, and raises rank-1 from 7.7 to 27.9% with 30% visible. So the ridge test could be relaxed for a deep matcher, while finger detection and framing remain necessary.

deep matcher, ridge-test rejectsEER% raw to fine-tuned ↓intervalEER gain on accepted images
30% visible29.6 to 23.6[-11.1, -0.9]13.0
50% visible23.6 to 17.6[-10.5, -2.6]7.2

The minutiae matcher also improves on these images (EER 46.9 to 42.2%, [-8.5, -0.2]). On the border rejects, mostly background or knuckles, and the low-fill rejects, mostly fingertips placed low in the crop (clean-crop rank-1 0 and 19% without the seven identities), completion changes little. We inspected all 152 images rejected for a tip at the border or low fill (Appendix D.2 covers MMFV).

Synthetic window shapes and occluded references

Unseen window shapes still help the deep matcher, while under the minutiae matcher the ellipse goes from 28.8 to 30.4% (Appendix D, Table 9). With references occluded like the probe, rank-1 is 44.8%, against 55.2% with no references (Table 10), so Table 1 describes captures with complete references.

unseen window, 30% visibledeep EER% raw to SweepFill ↓
ellipse30.2 to 16.7
blob38.5 to 19.9

Duplicate identities and the rescore

The result holds without the seven identities whose probes or references repeat enrollment imagery (four that fall back to enrollment crops as references, one of them also with a probe identical to a gallery image, and three with references identical to their enrollment images). Appendix B now rescores without these seven identities, and Table 37 keeps all 92 identities. With 30% visible the zero-shot model then lowers the deep matcher's EER from 28.8 to 17.7%, and the fine-tuned model lowers it to 15.8% and the minutiae EER from 39.9 to 29.2%.

30% visible, EER% ↓88 identities (four with enrollment crops as references removed, submitted)85 identities (seven removed, now)
raw probe, deep matcher28.528.8
raw probe, minutiae matcher40.139.9
fine-tuned model, minutiae matcher29.629.2

Appendix B now states the reasons and the effect on applicability.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A brief summary of the discussion for the Area, Program and Senior Area Chairs

We thank the Area, Program and Senior Area Chairs for overseeing the review of our paper, and the reviewers for their constructive feedback. Reviewers iEzK and 5Qto confirmed that our replies address their main concerns. We hope our reply to Reviewer WSPN, whose questions led to the subject-level fusion tests and the comparison with TransRef, addresses their concerns as well. For convenience, we briefly summarize the main concerns of each reviewer and how we resolved them.

Reviewer iEzK

  • Is the full model evaluated on subjects not used for model selection? On 14 subjects never used for any choice, completion lowers EER by 13.6 and 9.5 points and raises rank-1 by 47.2 and 49.0 points with 30% of the print visible, all significant over subjects.
  • What explains the gain over LaMa? At the larger training budget, SweepFill leads LaMa by 31.2 rank-1 points, and the reference frames account for 22.9 of them.
  • Real partial captures and the ISPFDv2 exclusions. We report every rejection reason by capture condition, and on 425 rejected images completion lowers EER from 35.6 to 28.0%. As the reviewer suggested, the final version will also clarify the scope of our results for naturally partial fingerprints.

Reviewer WSPN

  • Value over the reference frames, with subject-level uncertainty. Adding the completed probe lowers the two frames' EER from 12.5 to 8.3% with 30% of the print visible, significant over subjects, and with 70% visible it raises rank-1 by 14.6 points.
  • Comparison with a reference-guided inpainter. Against TransRef trained on the same inputs and budgets, SweepFill has the lower hidden-region orientation error (10.9° against 12.6°) and a higher rank-1 on every one of three training seeds (79.5% against 70.8% on average).

Reviewer 5Qto

  • Readability. The main text now uses shorter sentences, and Tables 1 and 2 carry the key comparisons. As the reviewer encouraged, the final version will improve readability further.
  • Diffusion models given the sweep. With 30% of the print visible, SweepFill reaches 79.2% rank-1, against 54.2% for DiffuEraser and 29.2% for LeftRefill trained on our crops.
  • Multi-frame fusion in the main table. Table 1 now includes fusion, and fusing the frames with the output of SweepFill gives the lowest EER among the degraded-frame options (8.3%).

The final version of the paper will include all of these analyses. We thank all three reviewers, iEzK, WSPN and 5Qto, again for their time and for suggestions that strengthened our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.