Apparent age estimation is a valuable tool for business personalization, yet current models frequently exhibit demographic biases. We review prior works on the DEX method by applying distribution learning techniques such as Mean-Variance Loss (MVL) and Adaptive Mean-Residue Loss (AMRL), and evaluate them in both accuracy and fairness. Using IMDB-WIKI, APPA-REAL, and FairFace, we demonstrate that while AMRL achieves state-of-the-art accuracy, trade-offs between precision and demographic equity persist. Despite clear age clustering in UMAP embeddings, our saliency maps indicate inconsistent feature focus across demographics, leading to significant performance degradation for Asian and African American populations. We argue that technical improvements alone are insufficient; accurate and fair apparent age estimation requires the integration of localized and diverse datasets, and strict adherence to fairness validation protocols.
Figures & tables
Figure 1. Examples of non-facial images in IMDB-WIKI. Note the presence of pixel-stretching artifacts.
Figure 2. APPA-REAL demographic distribution. The Caucasian group is highly over-represented.
Figure 3. IMDB-WIKI and APPA-REAL age distributions
Figure 4. FairFace demographic distribution
( Rothe et al., 2015 )
( Pan et al., 2018 )
( Zhao et al., 2022 )
CLAP 2015
FG-NET, CLAP 2016
FG-NET, CLAP 2016
MAE
ϵ
MAE
ϵ
MAE
ϵ
DEX
3.22
0.28
3.09
-
4.63
-
MVL
-
-
2.68
0.29
3.95
0.40
AMRL
-
-
-
-
3.61
0.39
Table 1. Comparison of DEX, MVL, and AMRL loss as reported by each paper
MAE
Testing set
Model
Appa.*
Real
ϵ -error
APPA-REAL
CLAP +
12.57
11.20
0.88
IMDB-WIKI + CLAP ∗
5.18
6.84
0.38
IMDB-WIKI ∗
6.96
7.89
0.49
IMDB-WIKI + APPA-REAL +
4.32
6.26
0.34
IMDB-WIKI + CLAP +
5.51
7.04
0.41
Table 2. MAE and ϵ -error results. These results do not yet include the FairFace finetuning step as we attempt to reproduce the results of ( Rothe et al., 2015 ) while comparing CLAP and APPA-REAL. Star symbol ( ∗ ) indicates Caffe model from ( Rothe et al., 2015 ) . Plus symbol ( + ) indicates our own implementation.
MAE
Model
Race
Gender
Apparent
Real
ϵ
AMRL: IMDB-WIKI + APPA-REAL
African American
Female
2.93
5.15
0.27
Male
2.92
6.45
0.23
Asian
Female
4.33
6.41
0.38
Male
3.77
4.92
0.36
Caucasian
Female
4.08
6.14
0.32
Table 3. MAE and ϵ -error results per model on APPA-REAL, grouped by race and gender, ordered by MAE (apparent age). We present results from the 5 best-performing models on apparent age. The MAE values for the best and worst groups per model are shown in bold.
Figure 5. Cosine similarity of each image with average embeddings for age, grouped by race
Table 4. MAE results per model ordered by MAE (apparent). We show the top 10 best-performing models in terms of apparent age MAE. Best model results are shown in bold.
Figure 6. UMAP-reduced embeddings on APPA-REAL test set using different finetuning combinations with cross-entropy loss. Left column, top to bottom: CLAP only, IMDB-WIKI + CLAP, and IMDB-WIKI + FairFace + CLAP. Right column replaces CLAP with APPA-REAL.
Figure 7. UMAP-reduced embeddings using different finetuning combinations with MVL (left column) and AMRL (right column). Follows the same progression of APPA-REAL-based models as in Figure 6 . AMRL with FairFace finetuning (bottom right) appears to result in stronger clustering of age ranges.
Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.
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