{"id":"66001483-c3c6-48f8-a7be-48a3da8d20c2","arxiv_id":"2412.13063","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A smartphone app with automated focus/zoom and on-device iris segmentation captures visible-spectrum iris images that match at 96.57% TAR (0.01% FAR) on a new 47-subject dataset.","lead":"This paper presents an Android app that uses automatic focus and zoom to capture high-quality visible-light iris images on a smartphone, plus a lightweight U-Net to segment the iris. On a new 47-subject dataset, the system reaches about 96.6% true acceptance at 0.01% false acceptance for visible-light iris matching, including matching against near-infrared enrollment images.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported TAR is computed in-sample: segmentation fine-tuned and FAR threshold set on the same 47 subjects, so the headline accuracy lacks out-of-sample support.","rationale":"The reader's weakest_assumption correctly identifies the small-sample, in-sample evaluation as the central threat to the headline accuracy. My reading agrees: the paper's strongest claim is the absolute TAR values, and those values depend on tuning both the segmentation model and the decision threshold on the very data used for testing. This is a concrete correctness risk, not a stylistic disagreement. The manuscript has additional editing issues (duplicated sections III.A and III.B in Section IV, an unresolved 'Figure ??' placeholder, and Fig. 9's 'TNormalization' caption), but those do not affect the core argument as much as the evaluation protocol. The internal consistency of the TARs across left/right eyes, distances, and colors is a modest positive signal, but it does not resolve the out-of-sample question. Because the reader already conditioned acceptance on these concerns, my stress-test does not move the verdict; it sharpens the required validation step: a subject-disjoint retraining and threshold-setting experiment with confidence intervals.","tokens_in":14741,"tokens_out":3043,"duration_ms":28485,"concrete_test":"Perform a subject-disjoint evaluation: split the 47 subjects into a development set (e.g., 30 subjects) and a test set (17 subjects). Fine-tune G-ATTU-Net only on the development subjects; set the Hamming-distance threshold from the impostor distribution of development subjects at FAR=0.01%; then compute TAR on the held-out test subjects' VIS and cross-spectral comparisons, with a 95% bootstrap confidence interval. If the held-out TAR falls below 90% or the CI width exceeds ±5%, the in-sample numbers in Section VIII are not reliable estimates of real-world accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the Section VIII evaluation estimates real-world generalization. The TAR of 96.57% (VIS), 97.95% (NIR), and ~96.2–96.3% (cross-spectral) is computed on the same 47-subject CUVIRIS set on which G-ATTU-Net was fine-tuned (Section III.B.1) and on which the decision threshold was set to hit FAR=0.01%. This creates a double in-sample advantage: segmentation masks for the test irises are produced by a model that has seen those irises during fine-tuning, and the threshold is placed using the impostor Hamming-distance distribution of the same data. With only 47 subjects, the impostor pool is small (roughly 10^4 comparisons per eye), so the 99.99th percentile used for FAR=0.01% is poorly determined; the TAR at that threshold can shift substantially with one or two impostor scores. Confidence intervals are not reported, so the reader cannot tell whether 96.57% versus, say, 90% is real. The Discussion (Section IX) acknowledges the small controlled dataset, but the abstract's definitive TAR numbers do not reflect this uncertainty. This is not an allegation of intentional bias; it is a structural limitation of evaluating a system on its training and threshold-tuning data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes a smartphone-based iris recognition system for visible-spectrum capture. It introduces an Android application that uses YOLOv3-tiny for eye/iris detection, automated zoom and focus adjustments, ISO/IEC 29794-6 quality checks, and a lightweight Ghost-Attention U-Net (G-ATTU-Net) for iris segmentation. The authors introduce the CUVIRIS dataset of 47 subjects with 940 VIS and NIR images, and report verification results in terms of TAR at FAR 0.01% for VIS, NIR, cross-spectral, distance-based, and iris-color-based comparisons. The central claim is that high-quality smartphone-captured VIS iris images can achieve recognition accuracy comparable to NIR, and that the proposed app and segmentation model make this practical on mobile devices.","tokens_in":14935,"tokens_out":5685,"duration_ms":53792,"significance":"If the reported performance held out-of-sample, this would be a useful contribution to mobile iris recognition: the CUVIRIS dataset is a novel public resource with modern smartphone captures and ISO-based quality checks, and the G-ATTU-Net model is a plausible lightweight alternative to heavier segmentation networks. The cross-spectral NIR-to-VIS comparison also addresses an application-relevant scenario. However, the current evaluation does not yet support the generalization claims in the abstract. The main strengths are the dataset, the real-time mobile capture pipeline, and the clear description of the app's control loop; the main weakness is that the accuracy numbers are computed in-sample, with the segmentation model fine-tuned on the same data used for evaluation and with the operating threshold set on the same small cohort. These issues are structural and require additional experiments or analysis rather than simple rewording.","major_comments":[{"comment":"The G-ATTU-Net segmentation model is fine-tuned on the CUVIRIS dataset (Section III.B.1) and then used to produce masks for the same CUVIRIS images in the recognition evaluation of Section VIII. No subject-disjoint or image-disjoint split is described for this fine-tuning. Because the same subjects, and very likely the same images, appear in both fine-tuning and evaluation, the reported TAR values estimate the performance of a segmentation module that has already seen the test irises. This can inflate accuracy if the model has overfit to subject-specific appearance or to the specific acquisition conditions of this single session. The authors should either use a strict subject-disjoint split for segmentation training versus recognition evaluation, or show that the fine-tuning has no effect on the masks used in the recognition experiment.","section":"Section III.B.1 and Section VIII"},{"comment":"The paper reports TAR at FAR 0.01% but does not describe how the Hamming-distance decision threshold is selected. If the threshold is derived from the impostor score distribution of the same 47-subject evaluation set, as is typical for ROC-based reporting, the operating point is chosen in-sample. With 47 subjects the impostor comparison pool is small (on the order of 10^4 comparisons per eye), so the 99.99th percentile of impostor scores is imprecisely estimated; one or two additional impostor scores can shift the threshold materially. The authors should report confidence intervals for the TAR at the chosen FAR, and either fix the threshold using a held-out set or describe an a priori threshold-setting protocol.","section":"Section VIII"},{"comment":"The contribution of the capture application is not isolated. The paper motivates the dedicated app through automatic zoom, focus, and quality control, but the recognition experiments in Section VIII compare only images captured by the full app against NIR images and against previously published datasets. There is no comparison with images captured by the same phone's stock camera application, or with the app's auto-zoom or quality checks disabled, using the same recognition pipeline. Without this baseline, it is not possible to determine whether the reported TAR is attributable to the app's capture features, to the phone's sensor, or to the processing pipeline. This is load-bearing for the paper's central claim that the application enables high-quality VIS iris capture.","section":"Section IV and Section VIII"},{"comment":"All TAR figures are based on 47 subjects and a single session per subject, with no confidence intervals or subject-level bootstrap. For example, the VIS left-eye genuine comparison uses roughly 329 comparisons (47 subjects x 7 verification images), so a TAR of 96.57% corresponds to about 12 false rejects; the binomial uncertainty is several percentage points. The abstract's precise numbers therefore overstate the certainty of the estimates. At minimum, the authors should report confidence intervals and, where possible, subject-level bootstrap resampling.","section":"Section VI and Section VIII"}],"minor_comments":[{"comment":"The UBIRIS-V1 dataset is cited as [11], but reference [11] in the bibliography is Edwards et al. on iris color; the correct Proença et al. reference for UBIRIS-V1 appears to be missing or misnumbered.","section":"Section II, Section III.A"},{"comment":"The text contains an unresolved placeholder 'see Figure ??' in the bullet list of application features; all figure references should be resolved before submission.","section":"Section I"},{"comment":"The Iris-Pupil Ratio metric is listed with a range of 0-100 and threshold >20 in Table II, while the text states its actual range is 9.58 to 121.30; these values should be reconciled.","section":"Section IV.A.3 and Table II"},{"comment":"These subsections duplicate, almost verbatim, the YOLOv3-tiny and G-ATTU-Net descriptions already given in Section III.A and Section III.B. Please consolidate to avoid repetition.","section":"Section IV.A.4 and Section IV.A.5"},{"comment":"The impostor comparison protocol is ambiguous: 'a randomly selected enrollment template from one subject was compared against all other subjects' images' does not specify whether this is done once per eye or repeatedly, and whether the random selection affects the reported TAR. Please clarify.","section":"Section VIII"},{"comment":"The text calls the iris-color comparisons 'ROC curves' while the figure caption says 'DET plots'; please use consistent terminology.","section":"Section VIII and Figure 12"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick read on arXiv:2412.13063. The app and dataset are worth something; the headline accuracy is not yet supported. The real contributions are the Android capture pipeline — automated zoom/focus feedback driven by YOLOv3-tiny detections, ISO/IEC 29794-6 quality gating, on-device G-ATTU-Net segmentation — and the CUVIRIS dataset itself: 47 subjects, 940 images, VIS plus NIR from a modern phone and IrisGuard scanner. That is genuinely new and useful. G-ATTU-Net is a modest Ghost-module/attention U-Net variant, not a conceptual leap, but the deployment story is plausible.\n\nThe soft spot is structural and load-bearing. G-ATTU-Net is fine-tuned on CUVIRIS (Section III.B.1) and then used to produce masks for the same CUVIRIS images in the Section VIII recognition evaluation. The decision threshold is set on that same data to hit 0.01% FAR. So both the segmentation and the operating point are in-sample. With 47 subjects the impostor pool is small, so the 99.99th-percentile threshold is poorly determined; one or two impostor scores can move it a lot. There are no confidence intervals, and there is no baseline against the phone's stock camera app, so the app-specific capture benefit is asserted, not demonstrated. The Discussion (Section IX) mentions the small controlled dataset, but the abstract's 96–98% TAR numbers do not carry that caveat.\n\nThere are also serious editing problems: the paper reprints the YOLOv3-tiny and G-ATTU-Net subsections verbatim in Section IV, and the intro has an unfilled Figure reference. That kind of thing is easy to fix but suggests the manuscript was rushed.\n\nWho is this for? Researchers working on VIS iris recognition and mobile biometric capture. They will want the CUVIRIS dataset, and they should treat the reported TAR as an upper bound until there is an out-of-sample evaluation with honest error bars. I would accept it for peer review — the dataset and system deserve close scrutiny — but I would make the authors add a held-out evaluation, report CIs, and compare against stock-camera capture before it is publishable.","headline":"A useful capture pipeline and a new public VIS iris dataset, but the headline TAR is computed in-sample and should not be trusted as a real-world accuracy estimate.","tokens_in":714,"tokens_out":2106,"would_cite":false,"duration_ms":42426,"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":"A dedicated Android capture application enables a modern smartphone to acquire visible-spectrum iris images whose recognition accuracy approaches infrared-based iris systems, with reported true acceptance rates above 96% across spectra…","keywords":["iris recognition","visible spectrum","smartphone biometrics","CUVIRIS dataset","G-ATTU-Net","YOLOv3-tiny","ISO/IEC 29794-6","cross-spectral matching"],"falsifier":"Run the identical capture-and-matching pipeline with the G-ATTU-Net weights frozen before any CUVIRIS images are used, the FAR threshold fixed in advance, and evaluation on a disjoint set of subjects; if the TAR at 0.01% FAR falls clearly below 96%, the paper's generalization claim is contradicted.","tokens_in":14479,"feed_emoji":"👁️","tokens_out":8875,"duration_ms":71925,"temperature":0.7,"pith_summary":"This paper argues that a modern smartphone, with a purpose-built capture application, can acquire visible-spectrum iris images good enough for reliable biometric recognition, including matching against templates enrolled with near-infrared scanners. To that end the authors built an Android app that detects the eye and iris with a compact YOLOv3-tiny model, automatically adjusts focus and zoom, applies ISO/IEC 29794-6 quality checks, and segments the iris on-device with a lightweight G-ATTU-Net. They collected a new dataset, CUVIRIS, with visible and infrared iris images from 47 subjects, and report true acceptance rates of 96.57% (visible), 97.95% (infrared), and 96.17–96.31% for infrared-to-visible cross-spectral matching at a false acceptance rate of 0.01%. If these figures hold beyond the study's small, in-sample evaluation, the practical consequence is that iris recognition could run on ordinary phone cameras without infrared hardware, using existing NIR enrollments for verification.","feed_headline":"Visible light iris scan on a phone hits 96.6%","feed_subtitle":"An Android app's auto focus and quality checks let phone iris images match infrared enrollments at about 96%.","key_machinery":"The load-bearing mechanism is the capture-and-feedback loop in the Android application: a YOLOv3-tiny detector (a compact 24-layer object detector) locates the eye and iris in each frame, the app scales those coordinates back to the sensor image and continuously adjusts camera zoom from the eye-box width and focal point from the iris coordinates, then crops a 640×480 eye image. The cropped frame must pass two quality gates — an OpenCV sharpness check (Laplacian and FFT) and the ISO/IEC 29794-6 metrics such as iris-pupil contrast, concentricity, margin adequacy, and usable iris area — before being saved. For recognition, the iris is segmented by G-ATTU-Net, a U-Net variant that replaces standard convolutions with Ghost Modules (which generate feature maps by combining a standard convolution with a depthwise separable convolution) and inserts attention blocks in the skip connections to keep the model at roughly 8 million parameters, about a quarter of the attention U-Net baseline. The segmented mask and red-channel image are normalized to 512×64 via the Hofbauer parameterization (rubber-sheet unwrapping with circular boundary detection), converted to a Gabor iris code by the OSIRIS v4.1 toolkit, and compared with masked Hamming distance including ±7 bit shifts for rotation alignment.","core_discovery":"The paper's central discovery claim is that high-resolution smartphone cameras, guided by a dedicated capture application, can produce visible-light iris images whose recognition accuracy approaches that of dedicated near-infrared imaging. Concretely, on the CUVIRIS dataset of 47 subjects captured with a Samsung Galaxy S21 Ultra, the authors report a true acceptance rate of 96.57% (left iris) and 96.81% (right iris) for visible-spectrum verification, 97.95% for NIR verification, and 96.17%/96.31% when NIR enrollment templates are verified against visible-spectrum probes, all at a false acceptance rate of 0.01%. Performance stayed above 96% across capture distances of 25 cm and 50 cm and across blue, brown, and gray iris colors. The authors attribute this outcome to the combination of automated focus and zoom on the iris, ISO/IEC 29794-6 quality gating, and an on-device segmentation model (G-ATTU-Net) that keeps the capture and matching pipeline free of the blur, reflection, and mis-segmentation artifacts that previously degraded VIS iris recognition.","pith_inferences":["An implication the paper leaves implicit: the same capture loop could be evaluated for other ocular biometrics, such as periocular or sclera recognition, where the automatic zoom and focus control would likely help as much as it does for the iris.","One extension we would suggest: the authors' in-sample evaluation could be stress-tested by reusing the exact app and matching pipeline on an independent cohort, which would clarify how much of the 96–98% TAR is due to the method versus the specific 47 subjects and tuned threshold.","A further consequence, if the cross-spectral numbers replicate: NIR-to-VIS matching could act as a fallback for law-enforcement or border databases built on infrared enrollments, so smartphone VIS captures could be checked against legacy NIR galleries.","Natural next step not in the paper: because the app already records color and reflection information in the VIS channel, the same quality-checked frames could support presentation-attack detection without additional capture hardware."],"forward_implications":["If the reported accuracy generalizes, iris recognition no longer requires specialized near-infrared cameras; a standard phone camera with the app could serve as the capture device.","The cross-spectral result implies that users enrolled on an existing NIR iris system could later verify on a smartphone in visible light, preserving prior enrollments.","The G-ATTU-Net's small parameter count and real-time speed suggest iris segmentation can run entirely on-device, keeping biometric templates local and reducing privacy exposure.","The ISO/IEC 29794-6 quality gating demonstrated here gives future mobile iris systems a concrete acceptance criterion that filters unusable frames before they enter the recognition pipeline.","CUVIRIS, once released, gives the research community a modern-smartphone VIS plus NIR benchmark to test new capture and matching methods on."],"supporting_citations":[{"why":"Supplies the masked Hamming-distance formula used for iris-code comparison.","marker":"[4]"},{"why":"Provides UBIRIS-V1 images used to train the YOLOv3-tiny eye detector and G-ATTU-Net.","marker":"[11]"},{"why":"Provides UBIRIS-V2 images used as additional training data for the detection and segmentation models.","marker":"[12]"},{"why":"Supplies the ISO/IEC 29794-6 quality metrics and thresholds the app enforces before saving an image.","marker":"[13]"},{"why":"Identifies the IrisGuard IG-AD100 as the device that captured the NIR images in CUVIRIS for NIR and cross-spectral tests.","marker":"[14]"},{"why":"Provides the prior visible-spectrum result (GMR 91.01% at FMR 0.01%) that this work compares against.","marker":"[15]"},{"why":"Supplies the YOLOv3-tiny architecture used for real-time eye and iris detection in the capture loop.","marker":"[18]"},{"why":"Supplies the Attention U-Net baseline whose parameter count and segmentation accuracy G-ATTU-Net is measured against.","marker":"[23]"},{"why":"Supplies the Hofbauer parameterization technique that normalizes iris and mask to 512×64 for matching.","marker":"[24]"},{"why":"Supplies the OSIRIS v4.1 toolkit that generates Gabor-based iris codes and performs matching.","marker":"[25]"}],"fun_headline_variants":["Phone iris recognition hits 96.6% using visible light","Visible-light iris scan on phone matches NIR accuracy within 1%","Smartphone app captures high-quality iris images in visible light","No infrared needed: phone camera does iris ID at 96%","G-ATTU-Net boosts phone iris recognition to 96.6% TAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that accuracy measured on the same 47 subjects used to fine-tune the segmentation and set the 0.01% FAR threshold predicts how the system will perform on people and conditions it has not seen.","fun_headline_variants_meta":{"raw":{"variants":["Phone iris recognition hits 96.6% using visible light","Visible-light iris scan on phone matches NIR accuracy within 1%","Smartphone app captures high-quality iris images in visible light","No infrared needed: phone camera does iris ID at 96%","G-ATTU-Net boosts phone iris recognition to 96.6% TAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3825,"prompt_tokens":1007,"completion_tokens":2818,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":2724}},"tokens_in":623,"tokens_out":2818,"duration_ms":21928,"temperature":1.0,"reasoning_tokens":2724,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:27:43.653995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical capture-and-matching pipeline with the G-ATTU-Net weights frozen before any CUVIRIS images are used, the FAR threshold fixed in advance, and evaluation on a disjoint set of subjects; if the TAR at 0.01% FAR falls clearly below 96%, the paper's generalization claim is contradicted.","supporting_citations":[{"cited_title":"How iris recognition works,","cited_arxiv_id":null,"evidence_quote":"Supplies the masked Hamming-distance formula used for iris-code comparison."},{"cited_title":"Quantita- tive measures of iris color using high resolution photographs,","cited_arxiv_id":null,"evidence_quote":"Provides UBIRIS-V1 images used to train the YOLOv3-tiny eye detector and G-ATTU-Net."},{"cited_title":"The ubiris. v2: A database of visible wavelength iris images captured on- the-move and at-a-distance,","cited_arxiv_id":null,"evidence_quote":"Provides UBIRIS-V2 images used as additional training data for the detection and segmentation models."},{"cited_title":"ISO/IEC 29794-6:2015 Information Technology - Biometric Sample Quality -Part 6: Iris Image Data,","cited_arxiv_id":null,"evidence_quote":"Supplies the ISO/IEC 29794-6 quality metrics and thresholds the app enforces before saving an image."},{"cited_title":"IrisGuard,","cited_arxiv_id":null,"evidence_quote":"Identifies the IrisGuard IG-AD100 as the device that captured the NIR images in CUVIRIS for NIR and cross-spectral tests."},{"cited_title":"Iris imaging in visible spectrum using white led,","cited_arxiv_id":null,"evidence_quote":"Provides the prior visible-spectrum result (GMR 91.01% at FMR 0.01%) that this work compares against."},{"cited_title":"Yolo v3-tiny: Object detection and recognition using one stage improved model,","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv3-tiny architecture used for real-time eye and iris detection in the capture loop."},{"cited_title":"Attention guided u-net for accurate iris segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the Attention U-Net baseline whose parameter count and segmentation accuracy G-ATTU-Net is measured against."},{"cited_title":"Exploiting superior cnn-based iris segmentation for better recognition accuracy,","cited_arxiv_id":null,"evidence_quote":"Supplies the Hofbauer parameterization technique that normalizes iris and mask to 512×64 for matching."},{"cited_title":"Osiris: An open source iris recognition software,","cited_arxiv_id":null,"evidence_quote":"Supplies the OSIRIS v4.1 toolkit that generates Gabor-based iris codes and performs matching."}],"review_version":1}