pith. sign in

arxiv: 2604.09125 · v1 · submitted 2026-04-10 · 💻 cs.CV

Few-Shot Personalized Age Estimation

Jakub Paplh\'am , Vojt\v{e}ch Franc , Artem Moroz This is my paper

Pith reviewed 2026-05-10 18:01 UTC · model grok-4.3

classification 💻 cs.CV
keywords personalized age estimationfew-shot learningage estimationOpenPAEface analysisneural processescomputer visionbiometrics
0
0 comments X

The pith

Personalizing age estimates with a few reference photos of the same person improves accuracy over global face-to-age mappings.

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

Standard age estimation models learn one mapping from appearance to age for all faces, but individuals age at different rates due to genetics, lifestyle, and health. When a small number of reference images with known ages for the same person are provided at test time, this identity-specific context can be used to adjust the estimate. The work introduces OpenPAE, the first open benchmark with strict protocols for N-shot personalized age estimation, along with a hierarchy of methods ranging from simple arithmetic offsets to a conditional attentive neural process. Experiments demonstrate that personalization yields consistent gains that go beyond domain adaptation alone and that nonlinear methods outperform linear and arithmetic baselines.

Core claim

Existing age estimation methods treat each face as an independent sample and learn a global mapping from appearance to age. This ignores that the mapping is identity-dependent because people age at different rates. When reference images of the same person with known ages are available, personalization exploits this context to improve estimates. The paper introduces OpenPAE as the first open benchmark for N-shot personalized age estimation with strict evaluation protocols and compares a hierarchy of baselines from arithmetic offset through closed-form Bayesian linear regression to a conditional attentive neural process, showing that personalization improves performance, the gains are not mere

What carries the argument

The conditional attentive neural process that takes reference images of the same identity as context to produce a personalized age estimate.

If this is right

  • Personalization consistently improves age estimation performance across methods.
  • The gains from using reference images exceed what can be explained by domain adaptation alone.
  • Nonlinear methods such as the conditional attentive neural process outperform simpler arithmetic offsets and linear regression.
  • The OpenPAE benchmark and protocols enable reproducible evaluation of few-shot personalized age estimation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar reference-based personalization could be applied to other variable face attributes such as expression intensity or health markers.
  • Systems could combine a global model with on-the-fly personalization only when references become available, reducing the need for references in every case.
  • The number of references required for reliable personalization without overfitting remains an open parameter that future work could map as a function of identity variability.

Load-bearing premise

Reference images of the same person with known ages are available at test time and the identity-dependent component of aging can be captured reliably from a small number of them without overfitting.

What would settle it

Measuring whether age estimation error decreases when reference images of the target identity are added versus when they are withheld, on a dataset where each identity has multiple held-out photos with ground-truth ages.

Figures

Figures reproduced from arXiv: 2604.09125 by Artem Moroz, Jakub Paplh\'am, Vojt\v{e}ch Franc.

Figure 1
Figure 1. Figure 1: Task overview. A global predictor (bottom) estimates age from a single face without considering individual aging rates. In personalized age estimation (top), the predictor is additionally conditioned on an unordered context set D of reference images with known ages from the same identity, allowing it to adapt to the individual. 2 Related Work 2.1 Age Estimation Age estimation from facial images has been st… view at source ↗
Figure 2
Figure 2. Figure 2: OpenPAE Results. Identity-balanced MAEid ↓ as a function of the maximum number of allowed references N. The horizontal line marks the global baseline (N=0). The dotted line shows the average number of references actually available per identity. −40 −20 0 20 40 Target−Reference Age [years] 0 10−1 P.D.F. (log scale) Age Difference Distribution −20 0 20 Target−Reference Age [years] −2 0 2 Improvement (∆MAE) ↑… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of Temporal Distance on Personalization. Left: [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
read the original abstract

Existing age estimation methods treat each face as an independent sample, learning a global mapping from appearance to age. This ignores a well-documented phenomenon: individuals age at different rates due to genetics, lifestyle, and health, making the mapping from face to age identity-dependent. When reference images of the same person with known ages are available, we can exploit this context to personalize the estimate. The only existing benchmark for this task (NIST FRVT) is closed-source and limited to a single reference image. In this work, we introduce OpenPAE, the first open benchmark for $N$-shot personalized age estimation with strict evaluation protocols. We establish a hierarchy of increasingly sophisticated baselines: from arithmetic offset, through closed-form Bayesian linear regression, to a conditional attentive neural process. Our experiments show that personalization consistently improves performance, that the gains are not merely domain adaptation, and that nonlinear methods significantly outperform simpler alternatives. We release all models, code, protocols, and evaluation splits.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper argues that standard age estimation ignores identity-dependent aging rates and proposes to personalize estimates using N reference images of the same person with known ages. It introduces the open OpenPAE benchmark with strict protocols, establishes a hierarchy of baselines (arithmetic offset, Bayesian linear regression, conditional attentive neural process), and reports that personalization yields consistent gains beyond domain adaptation while nonlinear methods outperform simpler ones. All code, models, and splits are released.

Significance. If the central experimental claims hold, the work is significant for providing the first open benchmark and reproducible protocols for few-shot personalized age estimation, along with a clear hierarchy of methods that isolate the value of identity-specific aging modeling. The release of data splits and code is a concrete strength that enables future comparisons.

major comments (1)
  1. [Experiments] Experiments section (and abstract claim): the assertion that 'the gains are not merely domain adaptation' is load-bearing but lacks an explicit control. All reported baselines use the known ages from the N references; a matched baseline that performs appearance-based adaptation on the same reference images while treating ages as unavailable (e.g., unsupervised feature alignment or pseudo-labeling) is needed to isolate personalization of aging rates from general domain adaptation to the test identity's appearance statistics. Without this, the reported improvements could be explained by better modeling of individual appearance rather than identity-dependent aging.
minor comments (2)
  1. [Abstract] Abstract: no quantitative numbers, error bars, dataset sizes, or ablation details are provided, making it difficult to assess the magnitude of the claimed gains.
  2. [Benchmark description] The paper should clarify the exact N values used in the few-shot protocols and whether the reference images are drawn from the same distribution as the test images.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address the major comment regarding the isolation of personalization effects from domain adaptation below.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (and abstract claim): the assertion that 'the gains are not merely domain adaptation' is load-bearing but lacks an explicit control. All reported baselines use the known ages from the N references; a matched baseline that performs appearance-based adaptation on the same reference images while treating ages as unavailable (e.g., unsupervised feature alignment or pseudo-labeling) is needed to isolate personalization of aging rates from general domain adaptation to the test identity's appearance statistics. Without this, the reported improvements could be explained by better modeling of individual appearance rather than identity-dependent aging.

    Authors: We acknowledge the referee's point that a matched control using the reference images for appearance-based adaptation without access to their age labels would provide a stronger isolation of the contribution from identity-specific aging rates. Our current results compare personalization methods (which leverage the known ages) against global models and existing domain adaptation baselines, with nonlinear methods showing larger gains. However, we agree that an explicit unsupervised adaptation baseline on the same N references (e.g., via feature alignment or pseudo-labeling without age supervision) would more rigorously rule out explanations based solely on modeling individual appearance statistics. We will add this control experiment to the revised manuscript, report the results in the experiments section, and update the abstract and discussion accordingly to better support the claim. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation or claims; empirical results on new benchmark are self-contained.

full rationale

The paper introduces the OpenPAE benchmark with released splits and protocols, then evaluates a hierarchy of baselines (arithmetic offset, Bayesian linear regression, conditional attentive neural process) on it. No equations, predictions, or derivations are present that reduce by construction to fitted inputs or self-citations. The claim that gains are not merely domain adaptation is an empirical observation from the experiments rather than a tautological or self-referential step. All load-bearing elements rely on external data splits and standard model comparisons, making the work self-contained against the new benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms or invented entities are detailed beyond standard supervised learning assumptions and the introduction of a neural-process variant as a baseline.

pith-pipeline@v0.9.0 · 5465 in / 1050 out tokens · 70750 ms · 2026-05-10T18:01:09.627206+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    EURASIP Journal on Image and Video Processing2016(1), 47 (Dec 2016) 4

    Abousaleh, F.S., Lim, T., Cheng, W.H., Yu, N.H., Hossain, M.A., Alhamid, M.F.: A novel comparative deep learning framework for facial age estimation. EURASIP Journal on Image and Video Processing2016(1), 47 (Dec 2016) 4

    work page 2016

  2. [2]

    The Lancet Digital Health7(6) (Jun 2025).https://doi.org/ 10.1016/j.landig.2025.03.0023

    Bontempi, D., Zalay, O., Bitterman, D., Birkbak, N., Shyr, D., Haugg, F., Qian, J., Roberts, H., Perni, S., Prudente, V., Pai, S., Dekker, A., Haibe-Kains, B., Guthier, C., Balboni, T., Warren, L., Krishan, M., Kann, B., Swanton, C., {De Ruysscher}, D., Mak, R., Aerts, H.: Faceage, a deep learning system to estimate biological age from face photographs to...

    work page doi:10.1016/j.landig.2025.03.0023 2025

  3. [3]

    In: Chinese Conference on Biometric Recognition (2012),https://api

    Cao, D., Lei, Z., Zhang, Z., Feng, J., Li, S.: Human age estimation using rank- ing svm. In: Chinese Conference on Biometric Recognition (2012),https://api. semanticscholar.org/CorpusID:112704784

    work page 2012

  4. [4]

    Masset, R

    Cao, W., Mirjalili, V., Raschka, S.: Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Let- ters140, 325–331 (2020).https://doi.org/https://doi.org/10.1016/j. patrec.2020.11.008,http://www.sciencedirect.com/science/article/pii/ S016786552030413X3

    work page doi:10.1016/j 2020

  5. [5]

    IEEE Transactions on Image Processing26(6), 2825–2838 (2017).https://doi.org/10.1109/TIP.2017.26899983

    Gao, B.B., Xing, C., Xie, C.W., Wu, J., Geng, X.: Deep label distribution learning with label ambiguity. IEEE Transactions on Image Processing26(6), 2825–2838 (2017).https://doi.org/10.1109/TIP.2017.26899983

    work page doi:10.1109/tip.2017.26899983 2017

  6. [6]

    In: Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI 2018)

    Gao, B.B., Zhou, H.Y., Wu, J., Geng, X.: Age estimation using expectation of label distribution learning. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI 2018). pp. xx–xx (2018) 3

    work page 2018

  7. [7]

    In: Dy, J., Krause, A

    Garnelo, M., Rosenbaum, D., Maddison, C., Ramalho, T., Saxton, D., Shanahan, M., Teh, Y.W., Rezende, D., Eslami, S.M.A.: Conditional neural processes. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Ma- chine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 1704–1713. PMLR (10–15 Jul 2018),https://proceedi...

    work page 2018

  8. [8]

    CreateSpace, United States, 3rd ed edn

    Gelman, A., Carlin, J., Stern, H., Dunson, D., Vehtari, A., Rubin, D.: Bayesian Data Analysis. CreateSpace, United States, 3rd ed edn. (2013) 8

    work page 2013

  9. [9]

    URL https: //doi.org/10.1109/TPAMI.2007.70733

    Geng, X., Zhou, Z.H., Smith-Miles, K.: Automatic age estimation based on fa- cial aging patterns. IEEE Trans. Pattern Anal. Mach. Intell.29(12), 2234–2240 (Dec 2007).https://doi.org/10.1109/TPAMI.2007.70733,https://doi.org/ 10.1109/TPAMI.2007.707333

    work page doi:10.1109/tpami.2007.70733 2007

  10. [10]

    Image Vis

    Georgopoulos, M., Panagakis, Y., Pantic, M.: Investigating bias in deep face anal- ysis: The kanface dataset and empirical study. Image Vis. Comput.102, 103954 (2020),https://api.semanticscholar.org/CorpusID:2186654645

    work page 2020

  11. [11]

    Hanaoka, K., Grother, P., Ngan, M.L., Yang, J., Quinn, G.W., Hom, A.: Face analysis technology evaluation: Age estimation and verification. NIST Intera- gency/Internal Report NIST IR 8525, National Institute of Standards and Tech- nology, Gaithersburg, MD (2024).https://doi.org/10.6028/NIST.IR.8525, https://doi.org/10.6028/NIST.IR.85251, 4, 10

    work page doi:10.6028/nist.ir.8525 2024

  12. [12]

    Symmetry 10, 385 (2018),https://api.semanticscholar.org/CorpusID:535282284 16 J

    Jeong, Y., Lee, S., Park, D.G., Park, K.H.: Accurate age estimation using multi- task siamese network-based deep metric learning for frontal face images. Symmetry 10, 385 (2018),https://api.semanticscholar.org/CorpusID:535282284 16 J. Paplhám et al

    work page 2018

  13. [13]

    In: International Conference on Learn- ing Representations (2019),https://openreview.net/forum?id=SkE6PjC9KX9

    Kim, H., Mnih, A., Schwarz, J., Garnelo, M., Eslami, A., Rosenbaum, D., Vinyals, O., Teh, Y.W.: Attentive neural processes. In: International Conference on Learn- ing Representations (2019),https://openreview.net/forum?id=SkE6PjC9KX9

    work page 2019

  14. [14]

    IEEE Transactions on Image Processing31, 4761–4775 (2022)

    Li, W., Lu, J., Wuerkaixi, A., Feng, J., Zhou, J.: Metaage: Meta-learning personal- ized age estimators. IEEE Transactions on Image Processing31, 4761–4775 (2022). https://doi.org/10.1109/TIP.2022.31880612, 4

    work page doi:10.1109/tip.2022.31880612 2022

  15. [15]

    2017 IEEE Con- ference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp

    Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: Agedb: The first manually collected, in-the-wild age database. 2017 IEEE Con- ference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 1997–2005 (2017),https://api.semanticscholar.org/CorpusID:17552574, 5

    work page 2017

  16. [16]

    In: Fifth Workshop on Meta- Learning at the Conference on Neural Information Processing Systems (2021), https://openreview.net/forum?id=h9yIMMjRoje9

    Müller, S., Hollmann, N., Arango, S.P., Grabocka, J., Hutter, F.: Transformers can do bayesian-inference by meta-learning on prior-data. In: Fifth Workshop on Meta- Learning at the Conference on Neural Information Processing Systems (2021), https://openreview.net/forum?id=h9yIMMjRoje9

    work page 2021

  17. [17]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 3

    Niu, Z., Zhou, M., Wang, L., Gao, X., Hua, G.: Ordinal regression with multi- ple output cnn for age estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) 3

    work page 2016

  18. [18]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 3

    Pan, H., Han, H., Shan, S., Chen, X.: Mean-variance loss for deep age estimation from a face. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) 3

    work page 2018

  19. [19]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2026) 5

    Paplham, J., Franc, V.: Photo dating by facial age aggregation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2026) 5

    work page 2026

  20. [20]

    7th International Conference on Automatic Face and Gesture Recognition (FGR06) pp

    Ricanek, K., Tesafaye, T.: Morph: a longitudinal image database of normal adult age-progression. 7th International Conference on Automatic Face and Gesture Recognition (FGR06) pp. 341–345 (2006),https://api.semanticscholar.org/ CorpusID:214409264, 5

    work page 2006

  21. [21]

    2015, in 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), 252–257, doi: 10.1109/ICCVW.2015.41

    Rothe, R., Timofte, R., Van Gool, L.: Dex: Deep expectation of apparent age from a single image. In: 2015 IEEE International Conference on Computer Vision Work- shop (ICCVW). pp. 252–257 (2015).https://doi.org/10.1109/ICCVW.2015.41 3

    work page doi:10.1109/iccvw.2015.41 2015

  22. [22]

    Sandhaus, R., Keller, Y.: Relative age estimation using face images (2025),https: //arxiv.org/abs/2502.048522, 4

    work page arXiv 2025

  23. [23]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022) 2, 4

    Shin, N.H., Lee, S.H., Kim, C.S.: Moving window regression: a novel approach to ordinal regression. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022) 2, 4

    work page 2022

  24. [24]

    American Journal of Physical Anthropology169(4), 678– 688 (2019).https://doi.org/https://doi.org/10.1002/ajpa.23878,https: //onlinelibrary.wiley.com/doi/abs/10.1002/ajpa.238783

    Windhager, S., Mitteroecker, P., Rupić, I., Lauc, T., Polašek, O., Schaefer, K.: Facial aging trajectories: A common shape pattern in male and female faces is dis- rupted after menopause. American Journal of Physical Anthropology169(4), 678– 688 (2019).https://doi.org/https://doi.org/10.1002/ajpa.23878,https: //onlinelibrary.wiley.com/doi/abs/10.1002/ajpa.238783

    work page doi:10.1002/ajpa.23878 2019

  25. [25]

    2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Zhang, Z., Song, Y., Qi, H.: Age progression/regression by conditional adversarial autoencoder. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4352–4360 (2017),https://api.semanticscholar.org/CorpusID: 8107084

    work page 2017

  26. [26]

    Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. pp. 18676–18688 (06 2022).https://doi.org/10.1109/CVPR52688.2022. 018147

    work page doi:10.1109/cvpr52688.2022 2022