摘要
由于基准测试流程中存在的不一致性，导致已发表的年龄估计方法结果可靠性不足，这给不同方法之间的比较带来了挑战。以往研究声称，过去十年间通过专用方法实现了持续的性能提升；然而，我们的研究结果对这些说法提出了质疑。本文揭示了当前评估协议中两个看似简单却长期存在的问题，并提出了相应的解决方案。我们对当前最先进的面部年龄估计方法进行了全面的对比分析。令人意外的是，我们发现不同方法之间的性能差异微乎其微，远不及其他因素的影响，如面部对齐质量、面部覆盖范围、图像分辨率、模型架构以及预训练所用数据量等。基于这些发现，我们提出以FaRL作为骨干模型，并在所有公开数据集上验证了其有效性。相关源代码及确切的数据划分已公开发布于GitHub。
代码仓库
paplhjak/facial-age-estimation-benchmark
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| age-estimation-on-afad | ResNet-50-OR-CNN | MAE: 3.16 |
| age-estimation-on-afad | ResNet-50-DLDL-v2 | MAE: 3.15 |
| age-estimation-on-afad | ResNet-50-DLDL | MAE: 3.14 |
| age-estimation-on-afad | ResNet-50-Unimodal-Concentrated | MAE: 3.20 |
| age-estimation-on-afad | ResNet-50-Cross-Entropy | MAE: 3.14 |
| age-estimation-on-afad | ResNet-50-Mean-Variance | MAE: 3.16 |
| age-estimation-on-afad | ResNet-50-Regression | MAE: 3.17 |
| age-estimation-on-afad | ResNet-50-SORD | MAE: 3.14 |
| age-estimation-on-afad | FaRL+MLP | MAE: 3.12 |
| age-estimation-on-agedb | ResNet-50-DLDL | MAE: 5.80 |
| age-estimation-on-agedb | ResNet-50-Unimodal-Concentrated | MAE: 5.90 |
| age-estimation-on-agedb | ResNet-50-DLDL-v2 | MAE: 5.80 |
| age-estimation-on-agedb | ResNet-50-SORD | MAE: 5.81 |
| age-estimation-on-agedb | ResNet-50-OR-CNN | MAE: 5.78 |
| age-estimation-on-agedb | FaRL+MLP | MAE: 5.64 |
| age-estimation-on-agedb | ResNet-50-Cross-Entropy | MAE: 5.81 |
| age-estimation-on-agedb | ResNet-50-Regression | MAE: 6.23 |
| age-estimation-on-agedb | ResNet-50-Mean-Variance | MAE: 5.85 |
| age-estimation-on-cacd | ResNet-50-Cross-Entropy | MAE: 3.96 |
| age-estimation-on-cacd | ResNet-50-Regression | MAE: 4.06 |
| age-estimation-on-cacd | ResNet-50-OR-CNN | MAE: 4.01 |
| age-estimation-on-cacd | ResNet-50-DLDL-v2 | MAE: 3.96 |
| age-estimation-on-cacd | ResNet-50-SORD | MAE: 3.96 |
| age-estimation-on-cacd | ResNet-50-Mean-Variance | MAE: 4.07 |
| age-estimation-on-cacd | ResNet-50-Unimodal-Concentrated | MAE: 4.10 |
| age-estimation-on-cacd | FaRL+MLP | MAE: 3.96 |
| age-estimation-on-cacd | ResNet-50-DLDL | MAE: 3.96 |
| age-estimation-on-chalearn-2016 | FaRL+MLP | MAE: 3.38 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-DLDL | MAE: 2.81 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-Cross-Entropy | MAE: 2.81 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-DLDL-v2 | MAE: 2.82 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-SORD | MAE: 2.81 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-Regression | MAE: 2.83 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-Mean-Variance | MAE: 2.83 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-OR-CNN | MAE: 2.83 |
| age-estimation-on-morph-album2-se-1 | FaRL+MLP | MAE: 3.04 |
| age-estimation-on-morph-album2-se-1 | ResNet-50-Unimodal-Concentrated | MAE: 2.78 |
| age-estimation-on-utkface | ResNet-50-OR-CNN | MAE: 4.40 |
| age-estimation-on-utkface | ResNet-50-Cross-Entropy | MAE: 4.38 |
| age-estimation-on-utkface | ResNet-50-SORD | MAE: 4.36 |
| age-estimation-on-utkface | ResNet-50-DLDL | MAE: 4.39 |
| age-estimation-on-utkface | ResNet-50-DLDL-v2 | MAE: 4.42 |
| age-estimation-on-utkface | FaRL+MLP | MAE: 3.87 |
| age-estimation-on-utkface | ResNet-50-Regression | MAE: 4.72 |
| age-estimation-on-utkface | ResNet-50-Unimodal-Concentrated | MAE: 4.47 |
| age-estimation-on-utkface | ResNet-50-Mean-Variance | MAE: 4.42 |