{"id":"40e91632-1e49-4652-bef1-3befb3970dc1","arxiv_id":"2505.05374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Periocular eye images outperform iris-only images for estimating age in children aged 4-16, with the best model reaching MAE 1.33 years and 83.82% binary age-group accuracy on a longitudinal NIR dataset.","lead":"The paper compares deep learning age estimates from children's iris versus full periocular eye images, using eight years of NIR images from 288 children. It finds periocular images give noticeably better age estimates (average error 1.33 years) and argues this supports on-device age checks in VR and child apps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Periocular-vs-iris comparison is confounded by differing input resolution and normalization; a matched-resolution control is needed before the comparative claim is accepted.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditional is right, but for a different primary reason. The reader singled out birth-year label granularity as the weakest assumption; that issue is acknowledged in the Discussion and affects both modalities symmetrically, so it does not threaten the periocular-versus-iris comparison as strongly as the representation confound. Sections III-A2 and III-A3 describe the preprocessing: eye images are 320x240 grayscale; iris images are 256x32 rubber-sheet strips plus mask. Table 7 then interprets all subsequent differences as modality effects. Because the two inputs have different pixel budgets, aspect ratios, and geometric normalizations, the observed gap can be explained without appealing to anatomical informativeness. The proposed control—retraining the eye model at the iris input size—would settle this. If the gap persists, the claim gains support; if it collapses, the headline conclusion is an artifact. This does not change the verdict (still CONDITIONAL), but it shifts the condition from 'add confidence intervals and fix counts' to 'demonstrate the comparison is not confounded before publication.'","tokens_in":18350,"tokens_out":5592,"duration_ms":64996,"concrete_test":"Retrain MobileNetV3-Large on periocular eye images that have been resized to the same 256x32 input dimensions used for the iris condition, keeping all other hyperparameters identical; then compare accuracy and MAE against the reported iris model on the same subject-exclusive test split. If the eye advantage shrinks or reverses, the modality comparison is explained by resolution and representation rather than by periocular anatomy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim—periocular eye images outperform normalized iris images (83.82% vs 71.00% accuracy, MAE 1.33 vs 2.32 for MobileNetV3, Table 7)—is confounded by the input representations, not just by anatomical region. Eye inputs are grayscale 320x240 images (76,800 pixels), while iris inputs are Daugman-normalized 256x32 strips (8,192 pixels) plus a binary mask. The two modalities differ in region, resolution, aspect ratio, and geometric normalization. Eyelid shape, scleral boundaries, pupil dilation, and skin texture are geometrically discarded by iris normalization, and the iris strip has roughly 9x fewer pixels. The paper's explanation that periocular context is 'richer' (Section IV-C) is a restatement of the confound rather than evidence. No control experiment holds pixel budget or geometric representation fixed. Consequently, the headline 'periocular models consistently outperform iris-based models' does not establish that periocular anatomy is more informative; it may only show that a 320x240 image with unnormalized geometry contains more usable information than a heavily downsampled rubber-sheet strip. The acknowledged label-granularity issue (birth-year subtraction) is real but affects both modalities symmetrically and is therefore less central to the paper's main novel finding.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents OcularAge, an empirical comparison of iris versus periocular (eye) images for estimating the age of pediatric subjects (4–16 years) using a longitudinal NIR dataset of 21,922 eye images from 288 children collected over eight years with two sensors. The authors train six CNN backbones (EfficientNet-B3, MobileNetV3, ResNet-50, DenseNet-121, ConvNeXt-Tiny, MobileViT-S) in a multi-task framework with age regression and binary age-group classification, using subject-exclusive splits. They report best periocular performance of MAE 1.33 years and 83.82% accuracy (MobileNetV3), outperforming iris-based models; additional experiments examine age-group-wise error, model confidence trends, cross-sensor generalization (IG-AD100 training to iCAM T10 testing), and deployment latency on GPU, Jetson, and Oculus Quest 2.","tokens_in":18623,"tokens_out":5357,"duration_ms":54391,"significance":"The work addresses an under-explored problem and provides a useful benchmark with subject-exclusive evaluation and realistic deployment profiling. The strongest contributions are the longitudinal dataset utilization, the multi-architecture comparison, and the end-to-end latency measurements on embedded/VR hardware. The comparative claim that periocular images are more informative than normalized iris images is, however, not yet established because the two input representations differ not only in anatomical region but also in resolution and geometric normalization. The absolute performance numbers are also affected by coarse integer age labels and single-run results. With these caveats addressed, the benchmark would be a valuable reference for child-focused biometric age estimation.","major_comments":[{"comment":"The iris versus eye comparison is confounded by input representation. Eye inputs are 320x240 grayscale images (76,800 pixels) with unnormalized geometry, while iris inputs are Daugman-normalized 256x32 strips (8,192 pixels) plus a binary mask. The observed gap (e.g., MobileNetV3: 83.82% vs 71.00%, MAE 1.33 vs 2.32) may reflect pixel budget and normalization rather than anatomical informativeness. The attribution to 'richer spatial and contextual information' in Section IV-C is a restatement of the confound. To support the central comparative claim, the authors should add a control experiment that holds pixel budget and geometric representation fixed, such as resizing the normalized iris strip to the same resolution as the eye image or cropping the eye image to the iris region without rubber-sheet normalization.","section":"Section III-A2, IV-C, Table 7"},{"comment":"The ground-truth ages are computed as capture year minus birth year, yielding integer labels with up to one year of error. The reported MAE of 1.33 years is therefore not a direct measure of error against true chronological age. The authors acknowledge the granularity issue in the Discussion ('age labels were based on birth year metadata, which introduces some granularity issues') but do not bound or correct for it. A quantitative analysis, such as reporting expected error under a uniform within-year distribution or using month-level metadata where available, is needed before the absolute MAE claim can be taken at face value.","section":"Section III-A2 and Discussion"},{"comment":"All results are single-run point estimates from one random seed. Several conclusions rest on small differences between models (e.g., MobileNetV3 and EfficientNet-B3 both report MAE 1.33 in Table 7). Without multiple seeds or bootstrap confidence intervals, the model rankings and the 'consistent outperformance' claim are not statistically supported. Please report means and standard deviations over at least three seeds, or provide equivalent uncertainty estimates, for the main result tables.","section":"Section III-C6 and Tables 4–10"},{"comment":"The reported image counts are internally inconsistent. The text states that the IG-AD100 contributed 16,285 images and the iCAM T10 5,014 images (Section III-A1); a later sentence says 16,885 eye images with 5,337 iCAM images; and Table 1 lists 16,585 IG-AD100 and 5,337 iCAM eye images. After preprocessing, the iris counts are 16,422 and 5,197. These discrepancies make the dataset description unreliable and must be reconciled before reproducibility can be assessed.","section":"Section III-A1 and Table 1"},{"comment":"ConvNeXt-Tiny achieves 45.23% binary accuracy with F1=0.00 and 0% ±1-year accuracy, which is at or below chance and indicates a training failure rather than genuine underperformance. This suggests a possible bug in the architecture adaptation (e.g., input stem modification) or optimization schedule. The authors should diagnose this failure or exclude the model from the benchmark; as reported, it distorts the claim of benchmarking six architectures.","section":"Section IV-A, Table 4"}],"minor_comments":[{"comment":"The deployment table is referred to as 'Table ??' in the text; this should read 'Table 11.'","section":"Section IV-G"},{"comment":"The sentence 'No explicit segmentation or masking was applied to periocular images' is followed later by the description of normalization statistics computed from the training set; consider clarifying that no anatomical masks are used for eye inputs, to avoid ambiguity about masking versus intensity normalization.","section":"Section III-A2"},{"comment":"The cross-sensor results show EfficientNet-B3 achieving higher accuracy on the unseen sensor (85.86%) than on the same sensor (79.98% in Table 9). This non-monotonic behavior deserves an explanation, as it may reflect different test-set compositions rather than true sensor-agnostic performance.","section":"Section IV-F, Tables 9–10"},{"comment":"The claim of 'first demonstration that reliable age estimation is feasible from children's ocular images' should be scoped more carefully, since prior iris-age studies on broader populations are cited; the uniqueness claim should be compared against those works in the discussion.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope as an applied biometrics/medical imaging study. The central finding is potentially publishable, but the confound in the modality comparison and the data/reporting inconsistencies need substantial work. The control experiment on matched resolution is essential; without it, the comparative claim does not establish the anatomical superiority of periocular over iris information. The label-granularity and single-seed issues also affect the headline numbers. I recommend major revision and a careful adversarial re-check of the dataset counts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is the first longitudinal pediatric ocular age-estimation benchmark I know of, and that alone makes it worth a serious look. The data are real: more than 21,000 NIR images from 288 children aged 4–16 across eight years, two sensors, and subject-exclusive splits. The authors run six architectures through the same multi-task protocol, report cross-sensor tests and VR latency numbers, and are honest about several limitations up front (birth-year age labels, grayscale NIR only). MobileNetV3 at 1.33 years MAE and 83.82% binary accuracy is a plausible headline for eye images, and the cross-sensor resilience of EfficientNet-B3 is genuinely interesting. Credit where due: this is a solid empirical contribution in a neglected area, and the subject-exclusive split is the right call.\n\nThe soft spots are real, though. The central comparison—periocular beats iris—is confounded. Eye inputs are 320x240 grayscale images (76,800 pixels); iris inputs are 256x32 Daugman-normalized strips (8,192 pixels) plus a mask. Those differ in region, resolution, aspect ratio, and geometric normalization all at once. So 'periocular context is richer' is a restatement of the confound, not evidence. A matched-resolution control, for instance iris crops at a comparable pixel budget or eye images downsampled to strip dimensions, is needed before the comparative claim holds. The stress-test note lands.\n\nAlso, all numbers are single-run point estimates; there are no confidence intervals or repeated-seed variance. The coarse birth-year labels (up to a year granularity) are acknowledged, and they affect both modalities symmetrically, so that is a secondary issue—but the headline 1.33 MAE is error against coarse labels, not exact chronological age. Internal counts disagree between the abstract, the prose, and Table 1 (16,285 vs 16,585 vs 16,422, for example), which is minor but sloppy. Dataset and code are unreleased, so the benchmark is currently not independently checkable. The abstract's 'reliable age estimation' overclaims what a single internal dataset and single-run experiment can support.\n\nWho gets value: people working on child-safe VR, pediatric soft biometrics, or ocular age estimation. The thinking is clear and the contribution is a useful baseline, not a settled result. I would send it to peer review with a request for the matched-resolution control, uncertainty estimates, an access path to the data or code, and toned-down abstract claims. If the authors fix the confound, this becomes a solid contribution; as is, it is a promising but partially unsubstantiated comparison.","headline":"Useful first pediatric ocular age-estimation benchmark, but the headline periocular-vs-iris comparison is confounded by input resolution and normalization, and the 'reliable' claim outruns the evidence.","tokens_in":19153,"tokens_out":2590,"would_cite":true,"duration_ms":27794,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A child's age is readable from near-infrared eye images with a mean error of 1.33 years, and periocular context beats the iris alone.","keywords":["pediatric age estimation","iris biometrics","periocular recognition","deep learning","multi-task learning","near-infrared imaging","sensor bias","virtual reality"],"falsifier":"Compute the same metrics using exact ages from birth dates and capture dates on this longitudinal dataset, or on a similarly sized pediatric ocular dataset. If the mean absolute error with exact labels is markedly larger than the year-only MAE (for example, above 2 years), then the reported accuracy largely reflects label quantization rather than true age estimation.","tokens_in":18171,"feed_emoji":"👁️","tokens_out":5002,"duration_ms":47633,"temperature":0.7,"pith_summary":"This paper asks whether a child's age can be estimated from an eye image and which portion of the eye carries the most age-related signal. Using a longitudinal set of more than 21,000 near-infrared ocular images from 288 children aged 4 to 16, collected over eight years with two different sensors, the authors train multi-task convolutional networks to jointly predict an age group (4–9 versus 10–16 years) and an exact age. They report that models trained on whole periocular eye images consistently outperform models trained on the normalized iris alone: the best model reaches a mean absolute error of 1.33 years and 83.82% binary age-group accuracy. The paper presents this as the first demonstration that reliable pediatric age estimation from ocular images is feasible, with practical value for privacy-preserving age checks in child-centric applications.","feed_headline":"Whole-eye images estimate children's ages within 1.3 years","feed_subtitle":"On 21,000 NIR images from 288 kids, periocular models hit 83.82% age-group accuracy and beat iris-only models.","key_machinery":"The engine of the comparison is a multi-task deep-learning setup in which each CNN has two output heads: a binary classifier for age group and a regressor for exact age, trained jointly with focal loss for classification and mean squared error for regression. The two input modalities are matched pairs: a grayscale full-eye image at 320×240 pixels and the iris region segmented and normalized to a 256×32 strip (using an open-source iris segmentation toolkit) with a binary occlusion mask concatenated as a second channel. Subject-exclusive splits, identical training configurations, and the same six backbones across both modalities make the eye-versus-iris difference attributable to input content rather than to architecture or data leakage.","core_discovery":"The central claim is that periocular context—eyelids, skin folds, sclera, and lash lines—carries more age-discriminative information for children than the iris texture pattern itself, and this holds across every architecture tested. MobileNetV3-Large on grayscale eye images achieves a MAE of 1.33 years and 83.82% age-group accuracy, versus 2.32 years and 71.00% for the same model on normalized iris strips. The paper further claims to establish the first systematic longitudinal benchmark comparing iris and periocular input for pediatric age estimation, showing that the best models generalize to an unseen sensor (EfficientNet-B3: 85.86% accuracy, 1.37 MAE in cross-sensor evaluation) and that a quantized MobileNetV3 runs in about 10 ms per image on a standalone VR headset, meeting real-time constraints.","pith_inferences":["The reported 1.33-year MAE is measured against year-granularity labels; with month-exact birth dates the true error could be somewhat larger, and re-benchmarking with exact ages would give a fairer headline number.","The same framework could likely be transferred to RGB images from the eye-tracking cameras already built into many VR headsets, removing the need for an added near-infrared module.","Fusing iris and periocular streams, or replacing the hard 10-year boundary with a soft or continuous transition, may push accuracy higher, especially in the 8–10 age range where confidence currently dips.","The moderate iris-only performance hints that even a supposedly stable biometric trait can encode auxiliary age signals; higher-resolution longitudinal iris capture could test whether that signal is genuine developmental biology or an artifact of normalization."],"forward_implications":["Eye-based models outperform iris-based models for every architecture tested, so periocular imagery is the better input modality for pediatric ocular age estimation.","The best eye-based models remain stable across developmental age bins, with the hardest groups at the extremes (ages 4–6 and 13–16) and the most predictable range in mid-childhood.","Model confidence drops around ages 8–10, near the group boundary, suggesting that practical systems should treat this transition zone with uncertainty-aware logic and escalate to human review when confidence is low.","Cross-sensor experiments show that models trained on one NIR camera transfer to a second unseen camera with only modest degradation for MobileNetV3 and EfficientNet-B3, supporting deployment in heterogeneous hardware environments.","A lightweight model runs in under 10 ms per image on a standalone VR headset, indicating that on-device, privacy-preserving age checks are technically plausible in real-time child-centric applications."],"supporting_citations":[{"why":"Provides the longitudinal pediatric iris dataset and its collection protocol, which is the basis for the images used in this study.","marker":"[19]"},{"why":"Extends the longitudinal iris dataset over up to six years, confirming the data's suitability for studying age-related changes in children.","marker":"[20]"},{"why":"Reports measurable age-related changes in ocular biometrics such as axial length in children, motivating the premise that ocular features encode age.","marker":"[10]"},{"why":"Shows age-dependent changes in choroidal thickness across subjects aged 4 to 23, supporting the idea that ocular anatomy carries developmental signals.","marker":"[11]"},{"why":"Demonstrates that natural aging produces measurable structural variation in the iris, providing the basis for expecting age information in iris texture.","marker":"[6]"},{"why":"Prior work on age prediction from iris features in adults, which this study extends to a pediatric longitudinal setting.","marker":"[8]"},{"why":"Shows that iris texture can distinguish older from younger adult subjects, grounding the hypothesis that iris features carry age cues.","marker":"[13]"},{"why":"Documents cross-sensor degradation in deep iris models, motivating the paper's sensor-generalization experiments.","marker":"[15]"},{"why":"Supplies the open-source iris segmentation and normalization pipeline used to produce the normalized iris inputs.","marker":"[33]"}],"fun_headline_variants":["Periocular beats iris for estimating kids' ages from eye images","Eye-region images reveal children's age within 1.3 years","Periocular context outperforms iris in pediatric age estimation","Kids' ages computable from eye images, periocular beats iris"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Ground-truth ages are computed as birth year minus capture year, producing integer labels that can be off by up to nearly a year, and the reported errors are measured against this coarse clock.","fun_headline_variants_meta":{"raw":{"variants":["Periocular beats iris for estimating kids' ages from eye images","Eye-region images reveal children's age within 1.3 years","Periocular context outperforms iris in pediatric age estimation","Kids' ages computable from eye images, periocular beats iris"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3315,"prompt_tokens":1027,"completion_tokens":2288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":2214}},"tokens_in":643,"tokens_out":2288,"duration_ms":15265,"temperature":1.0,"reasoning_tokens":2214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:05:46.738806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the same metrics using exact ages from birth dates and capture dates on this longitudinal dataset, or on a similarly sized pediatric ocular dataset. If the mean absolute error with exact labels is markedly larger than the year-only MAE (for example, above 2 years), then the reported accuracy largely reflects label quantization rather than true age estimation.","supporting_citations":[{"cited_title":"Iris recognition performance in children: A longitudinal study,","cited_arxiv_id":null,"evidence_quote":"Provides the longitudinal pediatric iris dataset and its collection protocol, which is the basis for the images used in this study."},{"cited_title":"Longitudinal performance of iris recognition in chil- dren: Time intervals up to six years,","cited_arxiv_id":null,"evidence_quote":"Extends the longitudinal iris dataset over up to six years, confirming the data's suitability for studying age-related changes in children."},{"cited_title":"Ocular biometrics as a function of age, gender, height, weight, and its association with spherical equivalent in children,","cited_arxiv_id":null,"evidence_quote":"Reports measurable age-related changes in ocular biometrics such as axial length in children, motivating the premise that ocular features encode age."},{"cited_title":"Association of choroidal thickness with eye growth: a cross- sectional study of individuals between 4 and 23 years,","cited_arxiv_id":null,"evidence_quote":"Shows age-dependent changes in choroidal thickness across subjects aged 4 to 23, supporting the idea that ocular anatomy carries developmental signals."},{"cited_title":"Analysis of physical ageing effects in iris biometrics,","cited_arxiv_id":null,"evidence_quote":"Demonstrates that natural aging produces measurable structural variation in the iris, providing the basis for expecting age information in iris texture."},{"cited_title":"Age prediction from iris biometrics,","cited_arxiv_id":null,"evidence_quote":"Prior work on age prediction from iris features in adults, which this study extends to a pediatric longitudinal setting."},{"cited_title":"The prediction of old and young subjects from iris texture,","cited_arxiv_id":null,"evidence_quote":"Shows that iris texture can distinguish older from younger adult subjects, grounding the hypothesis that iris features carry age cues."},{"cited_title":"Deepirisnet: Deep iris representation with applications in iris recognition and cross-sensor iris recognition,","cited_arxiv_id":null,"evidence_quote":"Documents cross-sensor degradation in deep iris models, motivating the paper's sensor-generalization experiments."},{"cited_title":"Osiris: An open source iris recognition software,","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source iris segmentation and normalization pipeline used to produce the normalized iris inputs."}],"review_version":1}