Relative Age Estimation Using Face Images

TL;DR

The paper tackles facial age estimation by introducing a differential age estimation framework that refines a Baseline Age Regressor (BAR) using age differences to a retrieved set of reference faces with known ages. It combines a retrieval pipeline, a Differential Age Regression (DAR) network, and an age-augmentation scheme with KDE-based error modeling, enabling iterative refinement of predictions. Empirical results on MORPH II and CACD under the Subject-Exclusive protocol show state-of-the-art MAE (2.47 on MORPH II; 5.27 on CACD), with ablations highlighting the importance of nearest-neighbor retrieval, KDE sampling, and a multi-task DAR loss. The work also analyzes biases across age, gender, and ethnicity, informing considerations for fairer, real-world deployment of age estimation systems.

Abstract

This work introduces a novel deep-learning approach for estimating age from a single facial image by refining an initial age estimate. The refinement leverages a reference face database of individuals with similar ages and appearances. We employ a network that estimates age differences between an input image and reference images with known ages, thus refining the initial estimate. Our method explicitly models age-dependent facial variations using differential regression, yielding improved accuracy compared to conventional absolute age estimation. Additionally, we introduce an age augmentation scheme that iteratively refines initial age estimates by modeling their error distribution during training. This iterative approach further enhances the initial estimates. Our approach surpasses existing methods, achieving state-of-the-art accuracy on the MORPH II and CACD datasets. Furthermore, we examine the biases inherent in contemporary state-of-the-art age estimation techniques.

Figures (13)

• Figure 1: Differential age estimation. The Baseline Age Regressor (BAR) estimates the age $\widehat{a}_{q}$ of the input image $\mathbf{x}_{q}$. $\widehat{\mathbf{x}}_{q}$ is the CNN embedding of $\mathbf{x}_{q}$ used with $\widehat{a}_{q}$ to retrieve the set of reference images that are of age $\widehat{a}_{q}$ and most visually similar to $\mathbf{x}_{q}$. The ages of the reference images are known. The Differential Age Regressor (DAR) estimates the age differences between $\mathbf{x}_{q}$ and the reference images and uses them to refine $\widehat{a}_{q}$.
• Figure 2: Differential age estimation in the training and test phases. In training time, the groundtruth age $a_{q}$ of the input face image $\mathbf{x}_{q}$, is augmented using Eq. \ref{['equ:train sampling']} and used to retrieve the reference set of images $\{\boldsymbol{x_{r}}\}_{1}^{R}$. In the test phase, a Baseline Age Regressor (BAR) estimates $\hat{a}_{q}$, the age of $\mathbf{x}_{q}$ that is used to retrieve $\{\boldsymbol{x_{r}}\}_{1}^{R}$.
• Figure 3: The Differential Age Regressor (DAR) network. The embeddings of the query image $\boldsymbol{x_{q}}$ and the reference images $\{\boldsymbol{x_{r}}\}_{1}^{R}$ are computed by a CNN, while the initial age estimate of the query image $\hat{a}_{q}$ and reference ages $\{\boldsymbol{a_{r}}\}_{1}^{R}$ are encoded by an embedding layer. The embeddings are concatenated to $\boldsymbol{{\hat{x}}_{q}}$ and $\{\boldsymbol{{\hat{x}}_{r}}\}_{1}^{R}$. The DAR network uses the embeddings to estimate the age differences $\{{d}_{r}\}_{1}^{R}$ and the weights per reference $\{{w}_{r}\}_{1}^{R}$. The resulting difference estimate $\hat{\Delta}$ is the weighted average $\hat{\Delta}=\sum_{r}{w}_{r}{d}_{r}$ added to $\hat{a}_{q}$ to compute the age estimate.
• Figure 4: Reference faces retrieval. (a) The age $a_{q}$ of $\boldsymbol{x_{q}}$ is used to retrieve $\{\widehat{\boldsymbol{x}}\boldsymbol{_{r}}\}_{1}^{K}$ the embeddings of the reference images of the same age. (b) $\widehat{\boldsymbol{x}}\boldsymbol{_{q}}$, the embedding of $\boldsymbol{x_{q}}$ is used to retrieve $\{\widehat{\boldsymbol{x}}\boldsymbol{_{r}}\}_{1}^{P},$$P\ll K,$ the $P$ embeddings in $\{\widehat{\boldsymbol{x}}\boldsymbol{_{r}}\}_{1}^{K}$ closest to $\widehat{\boldsymbol{x}}\boldsymbol{_{q}}$. (c) We randomly sample $R<P$ images for the final reference set $\{\boldsymbol{x_{r}\}}_{1}^{R}$.
• Figure 5: Differential Age Estimation. This network estimates the age differential between the query image $\hat{\mathbf{x}}_{q}$ and a single reference image $\hat{\mathbf{x}}.$ Their embeddings are concatenated as pairs $\left\{ \hat{\mathbf{x}}_{q},\hat{\mathbf{x}}_{r}^{i}\right\} .$ A set of regressors $\left\{ R_{c}\right\} _{-C}^{C}$ estimates $\left\{ d_{c}\right\} _{-C}^{C}:$ the second-order age differentials around$\left\{ c\right\} _{-C}^{C}$. $\left\{ d_{c}\right\} _{-C}^{C}$ are weighed by $\left\{ w_{c}\right\} _{-C}^{C}$ computed by $FC_{W}$ and a Softmax layer as in Eq. \ref{['equ:diff estimate']}.