Few-Shot Personalized Age Estimation

Jakub Paplhám; Vojtěch Franc; Artem Moroz

cs.CVApr 10, 2026

Few-Shot Personalized Age Estimation

Authors: Jakub Paplhám, Vojtěch Franc, Artem Moroz

Organizations: Department of Cybernetics, Faculty of Electrical Engineering, CTU in Prague · Czech Institute of Informatics, Robotics and Cybernetics, CTU in Prague

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.

Figures & tables

Panel of 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 $\mathcal{D}$ of reference images with known ages from the same identity, allowing it to adapt to the individual.

Dataset	Train IDs	Med.	Mean	Max	Train Imgs	Age Range
		Images / ID
CSFD	41,060	7	33	1925	1,189,660	[0, 110]
	Test IDs				Test Tasks
CSFD	4,929	11	40	1286	200,065	[0, 102]
MORPH	13,158	3	4	52	3 $\times$ 54,628	[16, 77]
AgeDB	566	24	29	138	3 $\times$ 16,481	[1, 101]

Table 1: Benchmark statistics. CSFD serves as the in-domain test set (Regime B); the remaining datasets evaluate cross-dataset generalization (Regime A).

Figure 2: OpenPAE Results. Identity-balanced

\text{MAE}_{\text{id}}\downarrow

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.

Figure 3: Impact of Temporal Distance on Personalization. Left: Probability density of the age gap between target and reference images. Right: Relative improvement (

\Delta\text{MAE}_{\text{id}}

) of the Pair-Avg

N=1

model over the unpersonalized global baseline. With the age difference

|y^{\mathrm{tgt}}-y^{\mathrm{ref}}|

below 10 years, the model consistently improves upon the global baseline on all datasets. When confronted with the heavy temporal tails of AgeDB or KANFace, these large age differences fall out-of-distribution, causing the personalized model to perform worse than the unconditioned global predictor.

	Method		Number of References ( $N$ )
	Method	$N=0$	$N=1$	$N=3$	$N=5$	$N=10$
In-Domain Evaluation (Regime B)
CSFD	Offset	3.81	4.01	3.47	3.36	3.29
	BLR		3.59	3.20	3.07	2.96
	Attn-G		2.74	2.47	2.40	2.35
	Pair-Avg		2.62	2.47	2.44	2.42

Table 2: OpenPAE Results. Identity-balanced mean absolute error (

\text{MAE}_{\text{id}}\downarrow

) across all evaluated datasets and methods.

N

indicates the number of reference images in the context set. The Global column (

N=0

) is the standard unpersonalized baseline.

	Method		Different ID (1)			Different ID (mix)			Same ID
		$N\!=0$	$N\!=1$	$N\!=5$	$N\!=10$	$N\!=1$	$N\!=5$	$N\!=10$	$N\!=1$	$N\!=5$	$N\!=10$
In-Domain Evaluation (Regime B)
CSFD	BLR	3.81	4.50	4.19	4.05	4.51	3.81	3.65	3.59	3.07	2.96
	Attn-G		3.27	3.08	2.96	3.11	2.80	2.75	2.74	2.40	2.35
	Attn-S		3.30	3.12	3.00	3.13	2.81	2.75	2.62	2.32	2.27
Cross-Dataset Generalization (Regime A)

Table 3: Personalization or Domain Adaptation? Table shows

\text{MAE}_{\text{id}}\downarrow

. Same ID : Standard evaluation. Different ID (1) : Each reference from one wrong identity, age-matched. Different ID (mix) : Each reference from a different wrong identity, age-matched. The different ID results represent domain adaptation without personalization.

Explore similar work

Toward Ethical Facial Age Estimation: A Generalized Zero-Shot Benchmark Without Training on Children's Data
May 28, 2026Caio Petrucci, Leo Sampaio Ferraz Ribeiro, Sandra AvilaZero-Shot Learning OOD Generalization
1Instituto de Computac¸˜ao, Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil · 2Instituto de Ciˆencias Matem´aticas e de Computac¸˜ao, Universidade de S˜ao Paulo (USP), S˜ao Carlos, Brazil
VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation
Mar 27, 2026Rakib Hossain Sajib, Md Kishor Morol, Rajan Das Gupta +2Large Vision-Language Models VLM Evaluation
Department of Computer Science and Engineering, Begum Rokeya University, Rangpur, Rangpur, Bangladesh · EliteLab.AI, Queens, New York, United States · Department of Computer Science, American International University - Bangladesh, Dhaka, Bangladesh
MobileAgeNet: Lightweight Facial Age Estimation for Mobile Deployment
Apr 18, 2026Arun Kumar, Aswathy Baiju, Radu Timofte +1On-Device Inference
Computer Vision Lab, CAIDAS & IFI, University of Würzburg, Germany

May 28, 2026J/K move · Enter open · S save

Toward Ethical Facial Age Estimation: A Generalized Zero-Shot Benchmark Without Training on Children's Data

Caio Petrucci, Leo Sampaio Ferraz Ribeiro, Sandra Avila

1Instituto de Computac¸˜ao, Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil · 2Instituto de Ciˆencias Matem´aticas e de Computac¸˜ao, Universidade de S˜ao Paulo (USP), S˜ao Carlos, Brazil

Zero-Shot Learning OOD Generalization

Open paper AskPDF arXiv

Age estimation from facial images typically relies on training data that includes images of minors, a practice that raises serious ethical, legal, and privacy concerns. In this work, we propose a generalized zero-shot benchmark for facial age estimation that explicitly excludes children's data during training while still assessing model performance on younger populations. We revisit six widely used datasets and introduce standardized splits with strict age-group separation: samples aged 18-59 for training, validation, and testing; samples under 18 reserved exclusively for zero-shot evaluation; and samples 60+ as an unseen validation set for model selection under distribution shift. For datasets with identity annotations, subject-exclusive splits prevent identity leakage and better reflect real-world deployment conditions. Evaluating nine state-of-the-art age estimation methods under this protocol reveals that all evaluated methods consistently fail to generalize to unseen age groups, suffering substantial performance degradation -- on average 46.4%, and up to 52.8% -- relative to the supervised baseline. Moreover, models do not simply degrade: they systematically anchor predictions for unseen ages to nearby seen classes, a manifestation of the well-known seen-class bias in generalized zero-shot learning. By formalizing age estimation without children's data as a generalized zero-shot benchmark on existing datasets, this work highlights a critical gap between current modeling practices and real-world ethical constraints. Our benchmark provides a principled basis for evaluating models under restricted data regimes and encourages the development of methods that are robust to distribution shift and aligned with responsible data use.

Abstract

Explore similar work

Toward Ethical Facial Age Estimation: A Generalized Zero-Shot Benchmark Without Training on Children's Data

VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation

MobileAgeNet: Lightweight Facial Age Estimation for Mobile Deployment