Research map / Image generation / Editing and style transfer
A Call to Reflect on Evaluation Practices for Age Estimation: Comparative Analysis of the State-of-the-Art and a Unified Benchmark
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
The paper shows that published facial age estimation gains stem from random splitting and pipeline confounds rather than specialized loss functions, proposes a subject-exclusive evaluation protocol and unified benchmark, and proposes a FaRL-ViT-B-16 + MLP baseline.
Problem
Published facial age estimation results are unreliable because of inconsistent evaluation: random splits leak subjects across train and test sets, and methods change multiple pipeline components at once, so gains cannot be attributed to specific loss functions.
Approach
Define a subject-exclusive evaluation protocol with fixed splits and isolate each pipeline component; compare loss functions with architecture, alignment, coverage, and resolution held fixed, finding that pretraining data, coverage, and resolution matter more than the loss, and use a FaRL backbone as a baseline.
Key results
- Under unified subject-exclusive splits with the loss fixed, OR-CNN, DLDL and Mean-Variance give significant gains over cross-entropy only with random initialization; with ImageNet or IMDB-WIKI pretraining none is significantly better.
- Input resolution 64 to 256 px reduces MORPH MAE from 3.33 to 2.81 and CLAP2016 MAE from 11.34 to 4.49 for ResNet-50.
- FaRL + MLP with frozen features obtains the best results on AgeDB, CLAP2016 and UTKFace, e.g. CLAP2016 MAE 4.34 versus 4.49 for ResNet-50 cross-entropy, but is worse on MORPH (MAE 4.67 versus 2.81).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FaRL features with an MLP instead of the ResNet-50 baseline | FaRL + MLP age estimator · AgeDB, AFAD, CACD2000, CLAP2016, MORPH, and UTKFace | MAE: AgeDB 5.81 -> 5.64; AFAD 3.14 -> 3.12; CACD2000 3.96 -> 3.96; CLAP2016 4.49 -> 3.38; MORPH 2.81 -> 3.04; UTKFace 4.38 -> | Mixed |
| Use specialized age-estimation loss functions instead of cross-entropy after pretraining (ablation) | ResNet-50 age estimators · Intra-dataset age estimation | MAE | No clear effect |
| Use OR-CNN, DLDL, or Mean-Variance loss instead of cross-entropy without pretraining (ablation) | ResNet-50 age estimators · Intra-dataset age estimation | MAE | Helped |
| Use 256 × 256 rather than 64 × 64 input resolution (ablation) | ResNet-50 age estimator · AgeDB | MAE: 8.43 -> 5.81 | Helped |
| Use 256 × 256 rather than 128 × 128 input resolution (ablation) | ResNet-50 age estimator · AgeDB | MAE: 6.90 -> 5.81 | Helped |
| Apply the RoI Tanh-polar input transformation (ablation) | ResNet-50 age estimator · AgeDB | MAE: 5.81 -> 5.93 | Hurt |
| Increase facial coverage from eyes-and-mouth to full-head coverage (ablation) | ResNet-50 age estimator · AgeDB | MAE: 6.06 -> 5.81 | Helped |
| Increase facial coverage from chin-and-forehead to full-head coverage (ablation) | ResNet-50 age estimator · AFAD and MORPH | MAE: AFAD 3.11 -> 3.14; MORPH 2.79 -> 2.81 | Mixed |
| Use full alignment rather than crop alignment (ablation) | ResNet-50 age estimator · AgeDB and CLAP2016 | MAE: AgeDB 5.93 -> 5.84; CLAP2016 4.68 -> 4.49 | Helped |
| Replace ResNet-50 with another backbone architecture (ablation) | ResNet-50 age estimator · AgeDB | MAE: ResNet-50 5.81; EfficientNet-B4 5.76; ViT-B-16 9.07; VGG-16 6.02 | Mixed |
| Use FaRL backbone features rather than a standard backbone | Age estimation model · Intra-dataset and cross-dataset evaluations | MAE | Helped |
Surprising findings
- Changing the loss function has a negligible effect once the pipeline is fixed, while facial coverage, resolution and backbone choice shift MAE more.
- Previously reported continual improvements over a decade on MORPH appear to stem from random splitting and tuning on the test set rather than loss functions.
Limitations
- Comparisons are restricted to methods that modify only the last layer or loss, excluding methods that change multiple components.
- FaRL features are frozen and the approach performs poorly on MORPH, attributed to distribution mismatch with LAION.
- Evaluation is limited to ResNet-50 as the main backbone for method comparisons.
Experimental setup
Datasets: AgeDB AFAD CACD2000 CLAP2016 FG-NET MORPH UTKFace IMDB-WIKI LAION
Benchmarks: AgeDB AFAD CACD2000 CLAP2016 FG-NET MORPH UTKFace
Metrics: MAE average rank (Friedman test and Nemenyi critical difference)
Baselines: Cross-Entropy Regression OR-CNN DLDL DLDL-v2 SORD Mean-Variance loss Unimodal
Builds on: FaRL DLDL OR-CNN Mean-Variance loss SORD
Closest papers by meaning
- From Cradle to Cane: A Two-Pass Framework for High-Fidelity Lifespan Face Aging (NeurIPS 2025 · similarity 0.87)
- DAA: A Delta Age AdaIN Operation for Age Estimation via Binary Code Transformer (CVPR 2023 · similarity 0.87)
- Deep Cost-Sensitive and Order-Preserving Feature Learning for Cross-Population Age Estimation (CVPR 2018 · similarity 0.87)
- Continuous Face Aging via Self-Estimated Residual Age Embedding (CVPR 2021 · similarity 0.86)
- Deep Regression Forests for Age Estimation (CVPR 2018 · similarity 0.86)
- Age Progression and Regression with Spatial Attention Modules (AAAI 2020 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.