Pith. sign in

REVIEW 4 major objections 4 minor 37 references

Underage Detection through a Multi-Task and MultiAge Approach for Screening Minors in Unconstrained Imagery

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a multi-task network with a frozen face encoder and four underage-threshold heads improves both age estimation and minor detection on unconstrained images, and introduces two benchmarks for evaluating that robustness.

desk verdict Useful benchmark paper with honest ablations; the gains are real but modest, and the benchmark construction is too intertwined with the method to fully separate. read the letter →

arxiv 2506.10689 v2 pith:A4UGQLKL submitted 2025-06-12 cs.CV

Christopher Gaul , Eduardo Fidalgo , Enrique Alegre , Rocío Alaiz Rodríguez , Eri Pérez Corral This is my paper
classification cs.CV
keywords ageestimationunderagedetectionmulti-tasklearningdistributionshiftclassimbalancefacialunconstrainedimagerybenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that underage detection in unconstrained facial images should be modeled as a set of binary age-threshold tasks rather than as a byproduct of age regression. It proposes a multi-task network in which a frozen FaRL vision-language backbone feeds one age-regression head and four underage heads (under 12, 15, 18, and 21 years), trained with age-balanced mini-batch resampling and an age gap that suppresses ambiguous near-threshold samples. The authors introduce two benchmarks—ASORES-39k, a cleaned restricted test, and ASWIFT-20k, a wild test of extreme poses, expressions, and low image quality—and report that their final model 'F' reduces mean absolute age error from 4.175 to 4.068 years on ASORES-39k and raises under-18 detection F2 from 0.801 to 0.857 at a 1% false-adult rate, while on ASWIFT-20k F2 rises from 0.742 to 0.833 at near-0.99 recall. A sympathetic reader would take the contribution to be a trainable-light recipe (under a million parameters) that keeps high recall of minors under distribution shift without fine-tuning the 87M-parameter backbone.

What carries the argument

The mechanism that carries the argument is the MultiAge head: a frozen FaRL vision transformer (87M parameters, producing a 512-dimensional face embedding) feeding one age-regression output and four sigmoid binary heads for thresholds 12, 15, 18, and 21 years. The heads share the same two-layer MLP representation, so the binary tasks steer the shared features toward the legally critical age range while the regression retains a global age signal. Two training devices are load-bearing: age-balanced mini-batch resampling, which partitions ages into 12 bins and samples each bin with probability inversely proportional to its occupancy, and an age gap, which omits from each binary loss the samples within a relative interval around the threshold (empirically $f=6/5$, yielding gaps such as 18+[-3,+3] years). The paper shows the age gap only works when the binary heads share the representation with the age regressor, because the gap samples still contribute to the regression loss and thus to the shared features. The weight-sharing MLP of the baseline (one matrix reused twice) is replaced by independent layers with a bottleneck, which the paper attributes to an increased rank constraint and task-specific projections.

What would settle it

Run the final model F on the 114 test images removed during cleaning plus the 238 removed validation images, and on the raw (uncleaned) 110k-image test set; if under-18 F2 at a 1% false-adult rate falls back toward the 0.801 age-only baseline, the benchmark cleaning rather than the multi-task head accounts for the reported gain. A complementary check is to have human annotators build an independent test set of minors with blur, occlusion, low contrast, and non-frontal poses and compare model F to the age-only baseline on it.

Watch

Extended reading notes

Core claim

The central discovery is that dedicating a small multi-task head to legal age thresholds, on top of a frozen FaRL image encoder, improves both age regression and underage detection under distribution shift compared with using the age estimate alone. The paper's final configuration, model F, uses an independent two-layer MLP with a 256-unit bottleneck, shares its features across the age regressor and four sigmoid heads for thresholds $T \in \{12,15,18,21\}$, and trains on a cleaned multi-source dataset that excludes CACD2000, with age-balanced resampling and a relative age gap ($[T/1.2, 1.2T]$ rounded, e.g. 18+[-3,+3] years) applied to the binary labels. On ASORES-39k it achieves MAE 4.068 years and under-18 F2 0.857 at a 1% false-adult rate; on ASWIFT-20k it achieves F2 0.833 while keeping recall of minors near 0.99. The paper also establishes two benchmarks for evaluating age models under real-world shift, and reports ablation evidence that removing the age gap collapses the single-task underage detector's recall.

Load-bearing premise

The evaluation assumes that the cleaned ASORES-39k and ASWIFT-20k test sets are unbiased, hard samples of real-world images, so that removing images flagged by a preliminary FaRL-based cleaning pass does not systematically hide the hardest cases the final system must catch.

Editorial extensions

If this is right

  • Minor-screening systems can be built with a frozen 87M-parameter backbone and a trainable head under a million parameters, keeping 0.99 recall of under-18 subjects at a 1% false-adult rate on the reported benchmarks.
  • The multi-threshold design extends naturally to age gates at 12, 15, 18, and 21, with under-12 F2 above 0.95 in both restricted and wild tests, covering the legal boundaries used by age-restricted services.
  • Age-balanced resampling is the effective imbalance remedy; the paper finds that alpha-reweighted focal loss and focusing parameter changes do not further improve accuracy once resampling is in place.
  • Training-data curation (removing the noisy CACD2000 set and cleaning label errors) improves both regression and detection, suggesting that dataset quality is as decisive as architecture for in-the-wild age estimation.
  • Robustness to pose and expression extremes holds up, while blur, low contrast, low saturation, and Black subjects remain the main failure modes, pointing to where additional training data is needed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A fair test of the claim would run model F on the 114 test images removed during cleaning plus the 238 removed validation images; if under-18 F2 at a 1% false-adult rate falls back toward the 0.801 age-only baseline, the benchmark cleaning rather than the multi-task head accounts for the reported gain.
  • ASWIFT-20k's selection relies on external pose and expression estimators (InsightFace and MorphSet) and hand-chosen thresholds; if those tools err more for children or particular ethnicities, the 'wild' set's composition and difficulty are partly an artifact of the selection tools.
  • The four binary heads output probabilities that should be monotone non-increasing in the threshold; a cheap consistency check on that monotonicity could serve as an input-level confidence signal to route borderline cases to human review—an extension the paper does not explore.
  • The reported demographic breakdown suggests a concrete testable remedy: augment training with synthetic degradation (blur, contrast, saturation) and targeted expression transfer, then measure whether Black-subject precision and low-contrast F2 improve without sacrificing recall.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a multi-task architecture for facial age estimation and underage detection: a frozen FaRL vision-language backbone feeding a compact MLP with one age-regression head and four binary underage heads (12, 15, 18, 21 years), trained with age-balanced mini-batch resampling, an alpha-reweighted focal loss variant, and an age-gap that removes ambiguous samples near thresholds. The authors introduce the Overall Underage Benchmark, a cleaned multi-source training/test corpus, and define two test sets: ASORES-39k (a restricted in-distribution test) and ASWIFT-20k (a deliberately challenging wild-shift test). The main empirical claims are that the proposed model 'F' reduces MAE on ASORES-39k from 4.175 to 4.068 years and improves under-18 detection F2 from 0.801 to 0.857 at a 1% false-adult rate, while nearly sustaining 0.99 recall on ASWIFT-20k. Ablations cover training-set composition, data cleaning, resampling, multi-task heads, loss parameters, MLP architecture, and fine-tuning, with comparisons to DEX, MiVOLO, and the authors' own reproduction of the FaRL+MLP baseline.

Significance. If the benchmark and evaluation protocol are valid, the paper's contributions are practically important: a simple, parameter-efficient multi-task model for CSEM screening that maintains high recall under pose, expression, and imaging-condition shifts, plus two test sets that could become useful community resources. The release of code and the large cleaned training corpus are additional assets. The ablation study is unusually thorough, and the comparisons to DEX and MiVOLO provide context. However, the central quantitative claims rest on the soundness of the benchmark construction and the evaluation protocol; the test-set threshold selection and model-dependent cleaning are load-bearing and need to be addressed before the reported gains can be taken at face value.

major comments (4)
  1. [5.3.2, Tables 5 and 6] The reported F2 and recall figures are obtained after fixing the discriminator threshold on the ASORES-39k test set such that the false-adult rate equals 0.01 (Section 5.3.2: "we fix the threshold of the underage discriminators as the point where the false overage detection rate is at 10^-2, i.e. at Recall=0.99 on ASORES-39k"). This is a form of test-set threshold selection: the operating point is chosen with access to the test labels, so the F2 scores reflect each model's calibration at a test-set-optimal point. This can overstate absolute F2 values and, more importantly, can confound model comparisons if the models are not equally calibrated. The authors should either select thresholds on a validation split, report threshold-free metrics such as area under the DET curve, or demonstrate that the relative ranking and the magnitude of the F2 improvement are insensitive to the threshold choice.
  2. [3.1] The test set is cleaned using a preliminary FaRL-based model with "slightly more tolerant settings" whose exact criteria are not specified; only the training/validation criteria are described in the footnote. Because the baseline (model 0) and the proposed models share the same frozen FaRL backbone, removing the 114 test images flagged by the preliminary model may preferentially delete cases that are hard for the FaRL feature family. This could differentially reduce the baseline's error and inflate the reported improvements (MAE 4.175 to 4.068; F2 0.801 to 0.857). The authors should specify the exact test-cleaning criteria and, more importantly, report the main comparisons on the uncleaned test set to demonstrate that the conclusions do not depend on this cleaning step.
  3. [5, Tables 5-8] The principal results are reported as point estimates without error bars, confidence intervals, or significance tests. The claimed improvements are small relative to typical seed-to-seed variability of deep models (e.g., MAE 4.175 vs. 4.068 years; F2 0.801 vs. 0.857). The authors should run each configuration with multiple seeds and report means and standard deviations, or at least show that the ordering of models is stable across seeds. This is especially important because the differences among models D, E, and F in Table 7 are of the same magnitude as the reported gain over the baseline.
  4. [3.2.3] The construction of ASWIFT-20k relies on hand-picked selection thresholds (lowest 8% in contrast and sharpness, lowest and highest 4% in brightness and saturation) and on external models (InsightFace for pose, MorphSet for expression), with additional ad-hoc choices such as adding AFAD and CASIA-Face-Africa at half the percentages. These thresholds and weighting decisions are not justified or validated, and the resulting composition is dominated by AFAD (4515), CASIA-Face-Africa (3375), and Dartmouth (3391), which correlates with age and ethnicity. The robustness claim would be substantially strengthened by a sensitivity analysis around these selection thresholds (e.g., re-running with different percentiles) and by presenting the main results separately for each selection criterion, not only for the union. Table 10 partially addresses this, but the selection thresholds themselves remain unexamined.
minor comments (4)
  1. [3.1, 3.2.1, Figure 2] There is a numerical inconsistency in the test-set sizes: Section 3.1 states that "114 out of 109089 test images were removed," but Figure 2(c) reports "Overall test (size: 110018)" and Section 3.2.1 says the overall test consists of 110k images. Please clarify whether the removed 114 images come from the 110k set or from a different pool, and reconcile the numbers.
  2. [4.3, Eq. (9)] The age gaps listed in Eq. (9) are inconsistent with the stated f=6/5 and "rounding towards zero": for T=15 the interval should be [12,18] (gap [-3,+3]) and for T=21 it should be [17,25] (gap [-4,+4]) if truncation toward zero is used. The listed values correspond to standard rounding of 12.5 to 13 and 17.5 to 18. Please specify the rounding rule precisely.
  3. [Abstract, 4.1] The abstract says the MLP "shares features" across heads, which is true, but the term is ambiguous in light of Section 4.1's discussion of the weight-sharing MLP (WS) versus the independent MLP (IND) used in model F. Please rephrase to avoid confusion between shared features and shared weights.
  4. [Tables 6 and 7 captions] The captions state that the top three values in each column are highlighted, but the highlighting is not visible in the provided text. Please ensure the final version renders the boldface/highlighting so the reader can verify the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the reported gains are empirical test measurements, and the benchmark-construction concerns are validity risks rather than construction-level circularity.

full rationale

This is an empirical paper with no first-principles derivation whose conclusions could reduce to its own inputs by construction. The central claim—Model F improves MAE from 4.175 to 4.068 years on ASORES-39k and F2 from 0.801 to 0.857 at the 1% false-adult operating point—is a measured performance difference on a held-out test set. The threshold is explicitly fixed so that the false-adult rate is 0.01, which makes Recall=0.99 on ASORES-39k a definitional operating point rather than a discovered result; the paper states this clearly in the Table 5 caption, so it is not a hidden prediction. The age-gap and resampling hyperparameters are examined through ablations; even if some choices were tuned on the test data, that is an overfitting/selection concern, not a logical reduction of the reported result to the fitted parameter. Self-citations to prior work by coauthors ([3], [21], [22], and [4]) appear in related work and in one external comparison benchmark, but none of these citations carries the load of the central argument, and the model is also compared against independent external systems (DEX, MiVOLO, Chaves et al.). The use of a preliminary model to clean label errors in the test set could bias the benchmark in favor of the model family, and the paper itself acknowledges demographic and image-condition failure modes; however, this is a data-validity and benchmark-construction concern, not an instance where an equation, fitted parameter, or self-citation forces the claimed outcome by definition. No circular step satisfying the required evidence standard was found.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The model's performance rests on several empirically chosen hyperparameters, and on assumptions about label accuracy and the reliability of external models used to build the benchmark. No new physical or conceptual entities are proposed.

free parameters (6)
  • Age gap scale factor f = 6/5
    Chosen empirically to set the width of the excluded interval around each age threshold; affects all binary heads.
  • Age group boundaries for resampling = 4,8,12,16,20,24,28,32,36,42,50 (12 groups)
    Chosen to partition the training age range; affects sampling probabilities.
  • Focal loss gamma = 2
    Set following Lin et al. [32]; not optimized in this paper.
  • Alpha reweighting parameter = None (alpha=1 for best models)
    They define a schedule but find it does not help; best results use alpha=1.
  • Label cleaning thresholds = 0.5%/2% etc.
    Thresholds for selecting images for manual inspection; influence which images are removed.
  • ASWIFT-20k selection thresholds = pose>45°, lowest 8% contrast/sharpness, 4% brightness/saturation, etc.
    Chosen to create a challenging test set; affects composition of ASWIFT-20k.
assumptions (3)
  • domain assumption Age labels in the public datasets are accurate after cleaning
    Relies on the cleaning procedure (Section 3.1) to remove label noise; if labels are still wrong, results are unreliable.
  • domain assumption FaRL backbone provides a sufficiently discriminative face representation for age
    The backbone is frozen and not fine-tuned; performance depends on this representation.
  • domain assumption ASWIFT-20k selection models (InsightFace, MorphSet) are accurate enough to select the intended hard cases
    Errors in pose/expression estimation could mislabel test subsets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Underage Detection through a Multi-Task and MultiAge Approach for Screening Minors in Unconstrained Imagery." pith.science (2026). https://pith.science/paper/A4UGQLKL

@misc{pith2026250610689,
  author       = {Pith},
  title        = {Pith review of: Underage Detection through a Multi-Task and MultiAge Approach for Screening Minors in Unconstrained Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A4UGQLKL}},
  note         = {Machine review of arXiv:2506.10689}
}
abstract

Accurate automatic screening of minors in unconstrained images requires models robust to distribution shift and resilient to the under-representation of children in public datasets. To address these issues, we propose a multi-task architecture with dedicated under/over-age discrimination tasks based on a frozen FaRL vision-language backbone joined with a compact two-layer MLP that shares features across one age-regression head and four binary underage heads (12, 15, 18, and 21 years). This design focuses on the legally critical age range while keeping the backbone frozen. Class imbalance is mitigated through an $\alpha$-reweighted focal loss and age-balanced mini-batch sampling, while an age gap removes ambiguous samples near thresholds. Evaluation is conducted on our new Overall Underage Benchmark (303k cleaned training images, 110k test images), defining both the "ASORES-39k" restricted overall test, which removes the noisiest domains, and the age estimation wild-shifts test "ASWIFT-20k" of 20k-images, stressing extreme poses ($>$45{\deg}), expressions, and low image quality to emulate real-world shifts. Trained on the cleaned overall set with resampling and age gap, our multiage model "F" reduces the mean absolute error on ASORES-39k from 4.175 y (age-only baseline) to 4.068 y and improves under-18 detection from F2 score of 0.801 to 0.857 at 1% false-adult rate. Under the ASWIFT-20k, the same configuration nearly sustains 0.99 recall while F2 rises from 0.742 to 0.833, demonstrating robustness to domain shift.

Figures

Figures reproduced from arXiv: 2506.10689 by the authors.

Figure 1
Figure 1. Sample face patches from the different age estimation benchmarks, as processed by the model during training, i.e., with the random flipping, scaling, and cropping described in Section 4. contains Asian subjects with low resolution, and the face crops are tight, which leads 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The age distribution of the training and test data. (a) Number of images for each source dataset [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Distribution of ASORES-39k and ASWIFT-20k over gender, ethnicity (FairFace [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Selection of facial expressions for ASWIFT-20k. (a) Distribution of the selected “extreme” ex [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Scheme of the model architecture: The face patch extracted from the image scene is divided into [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Under/over 18-years discrimination accuracy obtained under different training schemes (cln: cleaned or not; CACD: with or without the CACD part of the training set; res: using resampling accord￾ing to Eq. 7 or not and with or without a dedicated under/over 18 discrimin…
Figure 8
Figure 8. Figure 8: Under/over 18-year discrimination accuracy. Variation points: use CACD2000 in the training or not; resampling (res) according to Eq. 7 or not; prediction head setup (Heads); reweighting (alpha) according to Eq. 6 or αy = 1. Common parameters: cleaned training sets (cln…
Figure 9
Figure 9. Figure 9: Error distribution by age and yaw angle on the ASWIFT-20k test. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    Angulu, J

    R. Angulu, J. R. Tapamo, A. O. Adewumi, Age estimation via face images: a survey, EURASIP JIVP (2018) 42. doi:10.1186/s13640-018-0278-6

    work page doi:10.1186/s13640-018-0278-6 2018

  2. [2]

    Eidinger, R

    E. Eidinger, R. Enbar, T. Hassner, Age and gender estimation of unfiltered faces, IEEE Transactions on information forensics and security 9 (2014) 2170–2179. doi:10.1109/TIFS.2014.2359646

    arXiv 2014

  3. [3]

    Gangwar, V

    A. Gangwar, V . González-Castro, E. Alegre, E. Fidalgo, A. Martínez-Mendoza, DeepHSAR: Semi-supervised fine-grained learning for multi-label human sexual activity recognition, Information Processing & Management 61 (2024) 103800. doi:10.1016/j.ipm.2024.103800. 23

    work page arXiv 2024

  4. [4]

    E. D. Jeuland, A. Del Río Ferreras, D. Chaves, E. Fidalgo, V . González-Castro, E. Alegre, Assessment of age estimation methods for forensic applications using non-occluded and synthetic occluded facial images, in: XLIII Jornadas de Au- tomática, Coruña, 2022, p. 972. doi:10.17979/spudc.9788497498418.0972

    work page arXiv 2022

  5. [5]

    Ricanek, T

    K. Ricanek, T. Tesafaye, MORPH: a longitudinal image database of normal adult age-progression, in: 7th International Conference on Automatic Face and Gesture Recognition (FGR06), 2006, pp. 341–345. doi:10.1109/FGR.2006.78

    doi:10.1109/fgr.2006.78 2006

  6. [6]

    Z. Niu, M. Zhou, L. Wang, X. Gao, G. Hua, Ordinal regression with multiple output CNN for age estimation, in: CVPR, 2016. URL

    work page 2016

  7. [7]

    Moschoglou, A

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia, S. Zafeiriou, AgeDB: the first manually collected, in-the-wild age database, in: CVPR Work- shops, 2017, pp. 51–59. URL

    work page 2017

  8. [8]

    Blanchard, A

    R. Blanchard, A. D. Lykins, D. Wherrett, M. E. Kuban, J. M. Cantor, T. Blak, R. Dickey, P. E. Klassen, Pedophilia, hebephilia, and the DSM-V, Archives of Sexual Behavior 38 (2009) 335–350. URL. doi:10.1007/s10508-008-9399-9

    work page doi:10.1007/s10508-008-9399-9 2009

Show all 37 references
  1. [9]

    Lanitis, C

    A. Lanitis, C. Taylor, T. Cootes, Toward automatic simulation of aging effects on face images, IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (2002) 442–455. doi:10.1109/34.993553

    2002 doi

  2. [10]

    H. Han, C. Otto, A. K. Jain, Age estimation from face images: Human vs. ma- chine performance, in: ICB, 2013. doi:10.1109/ICB.2013.6613022

    2013

  3. [11]

    LeCun, Y

    Y . LeCun, Y . Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444. doi:10.1038/nature14539

    2015 doi

  4. [12]

    G. Levi, T. Hassner, Age and gender classification using convolutional neural networks, in: CVPR Workshops, 2015. URL

    2015

  5. [13]

    Rothe, R

    R. Rothe, R. Timofte, L. V . Gool, Deep expectation of real and apparent age from a single image without facial landmarks, International Journal of Computer Vision 126 (2018) 144–157. doi:10.1007/s11263-016-0940-3

    2018 doi

  6. [14]

    Yang, Y .-H

    T.-Y . Yang, Y .-H. Huang, Y .-Y . Lin, P.-C. Hsiu, Y .-Y . Chuang, SSR-Net: A com- pact soft stagewise regression network for age estimation, in: IJCAI, 2018, pp. 1078–1084. doi:10.24963/ijcai.2018/150

    2018 doi

  7. [15]

    Q. Li, J. Wang, Z. Yao, Y . Li, P. Yang, J. Yan, C. Wang, S. Pu, Unimodal- concentrated loss: Fully adaptive label distribution learning for ordinal regres- sion, in: CVPR, 2022, pp. 20513–20522. URL

    2022

  8. [16]

    Paplhám, V

    J. Paplhám, V . Franc, A call to reflect on evaluation practices for age estimation: Comparative analysis of the state-of-the-art and a unified benchmark, in: CVPR, 2024, pp. 1196–1205. URL. 24

    2024

  9. [17]

    Zheng, H

    Y . Zheng, H. Yang, T. Zhang, J. Bao, D. Chen, Y . Huang, L. Yuan, D. Chen, M. Zeng, F. Wen, General facial representation learning in a visual-linguistic manner, in: CVPR, 2022, pp. 18697–18709. URL

    2022

  10. [18]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning transferable visual models from natural language supervision, volume 139 ofProceedings of Machine Learning Research, PMLR, 2021, pp....

    2021

  11. [19]

    Castrillón-Santana, J

    M. Castrillón-Santana, J. Lorenzo-Navarro, C. M. Travieso-González, D. Freire- Obregón, J. B. Alonso-Hernández, Evaluation of local descriptors and CNNs for non-adult detection in visual content, Pattern Recognition Letters 113 (2018) 10–18. doi:10.1016/j.patrec.2017.03.016

    2018 doi

  12. [20]

    Anda, N.-A

    F. Anda, N.-A. Le-Khac, M. Scanlon, DeepUAge: Improving underage age es- timation accuracy to aid CSEM investigation, Forensic Science International: Digital Investigation 32 (2020) 300921. doi:10.1016/j.fsidi.2020.300921

    2020

  13. [21]

    Chaves, E

    D. Chaves, E. Fidalgo, E. Alegre, F. Jáñez-Martino, R. Biswas, Improving age estimation in minors and young adults with occluded faces to fight against child sexual exploitation, in: Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Compute...

    2020 doi

  14. [22]

    Gangwar, V

    A. Gangwar, V . González-Castro, E. Alegre, E. Fidalgo, AttM-CNN: At- tention and metric learning based CNN for pornography, age and child sex- ual abuse (CSA) detection in images, Neurocomputing 445 (2021) 81–104. doi:10.1016/j.neucom.2021.02.056

    2021 doi

  15. [23]

    Agustsson, R

    E. Agustsson, R. Timofte, S. Escalera, X. Baro, I. Guyon, R. Rothe, Apparent and real age estimation in still images with deep residual regressors on appa-real database, in: FG, 2017. doi:10.1109/FG.2017.20

    2017 doi

  16. [24]

    Chen, C.-S

    B.-C. Chen, C.-S. Chen, W. H. Hsu, Cross-age reference coding for age-invariant face recognition and retrieval, in: Proceedings of the European Conference on Computer Vision (ECCV), 2014. doi:10.1007/978-3-319-10599-4_49

    2014 doi

  17. [25]

    Zhang, Y

    Z. Zhang, Y . Song, H. Qi, Age progression/regression by conditional adversarial autoencoder, in: CVPR, IEEE, 2017. doi:10.48550/arXiv.1702.08423

    2017 doi arXiv

  18. [26]

    K. A. Dalrymple, J. Gomez, B. Duchaine, The Dartmouth database of children’s faces: Acquisition and validation of a new face stimulus set, PLoS ONE (2013). doi:10.1371/journal.pone.0079131

    2013 doi

  19. [27]

    Muhammad, Y

    J. Muhammad, Y . Wang, C. Wang, K. Zhang, Z. Sun, CASIA-Face-Africa: A large-scale African face image database, IEEE Trans. Inf. Forensics Secur. 16 (2021) 3634–3646. doi:10.1109/TIFS.2021.3080496. 25

    2021

  20. [28]

    Karkkainen, J

    K. Karkkainen, J. Joo, FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation, in: IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV), 2021, pp. 1548–1558. URL

    2021

  21. [29]

    V onikakis, S

    V . V onikakis, S. Winkler, Efficient facial expression analysis for dimensional affect recognition using geometric features, arXiv:2106.07817 (2021). URL

    2021 arXiv

  22. [30]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Repr...

    2021

  23. [31]

    Caruana, Multitask learning, Machine Learning 28 (1997) 41–75

    R. Caruana, Multitask learning, Machine Learning 28 (1997) 41–75. doi:10. 1023/A:1007379606734

    1997

  24. [32]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, P. Dollar, Focal loss for dense object detection, in: ICCV , 2017. URL

    2017

  25. [33]

    J. An, L. Ying, Y . Zhu, Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients, in: ICLR, 2021. URL

    2021

  26. [34]

    X. Geng, Q. Wang, Y . Xia, Facial age estimation by adaptive label distribution learning, in: 2014 22nd International Conference on Pattern Recognition, 2014, pp. 4465–4470. doi:10.1109/ICPR.2014.764

    2014 doi

  27. [35]

    Rothe, R

    R. Rothe, R. Timofte, L. V . Gool, DEX: Deep expectation of apparent age from a single image, in: IEEE International Conference on Computer Vision Workshops (ICCVW), 2015. URL

    2015

  28. [36]

    Kuprashevich, I

    M. Kuprashevich, I. Tolstykh, Mivolo: Multi-input transformer for age and gen- der estimation, in: D. I. Ignatov et al. (Ed.), Analysis of Images, Social Networks and Texts, Springer Nature, 2024. doi:10.1007/978-3-031-54534-4_15

    2024 doi

  29. [37]

    Merler, N

    M. Merler, N. K. Ratha, R. S. Feris, J. R. Smith, Diversity in faces, CoRR abs/1901.10436 (2019). URL.arXiv:1901.10436. 26

    2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.