Multiview Benchmark of Action Unit Detection with the RIKEN Facial Expression Database

For this study, we benchmark four open-source facial action estimators (OpenFace 3.0, OpenFace 2.2.0, Py-Feat, and PyAFAR) on the RIKEN facial expression database (N = 4,800 images), which is annotated by facial intensity (0–5). Using multiple metrics, we evaluated their detection performance. When using the subset of facial actions common across systems, the average Area Under the Curve scores of the three modern systems were comparable (OpenFace 3.0 = .82; Py-Feat = .83; PyAFAR = .82) and higher than OpenFace 2.2.0 (.75). Facial component-wise analyses revealed systematic differences: OpenFace 3.0 performs best for eyebrow raises and mouth opening; OpenFace 2.2.0 better captures frown/tension-related movements; and PyAFAR excels at smile-related facial action units. These findings, which indicate that optimal tool choice is facial-component-dependent, underscore the value of per-action units.