Table 3.

AU detection performances on CLOCK dataset

−	S	AUC	PA	NA
(a) AU4
Study 1: CLOCK $\to$ CLOCK	0.82	0.79	0.61	0.96
Study 2: (MIAMI + CLOCK) $\to$ CLOCK	0.78	0.80	0.57	0.94
Study 3: MIAMI $\to$ CLOCK	0.81	0.60	0.31	0.95
Hammal et al. (2017)	0.84	−	0.19	0.96
OpenFace	−0.01	0.62	0.26	0.62
Study 4 - Adult AFAR: (EB+ + GFT) $\to$ CLOCK	0.70	0.72	0.46	0.91
Study 5 - Infant AFAR: (EB+ + GFT + MIAMI + CLOCK) $\to$ CLOCK	0.77	0.79	0.56	0.93
(b) AU6
Study 1: CLOCK $\to$ CLOCK	0.73	0.87	0.81	0.89
Study 2: (MIAMI + CLOCK) $\to$ CLOCK	0.75	0.88	0.83	0.90
Study 3: MIAMI $\to$ CLOCK	0.68	0.78	0.72	0.89
Hammal et al. (2017)	0.74	−	0.76	0.91
OpenFace	0.65	0.83	0.69	0.89
Study 4 - Adult AFAR: (EB+ + GFT) $\to$ CLOCK	0.67	0.77	0.69	0.89
Study 5 - Infant AFAR: (EB+ + GFT + MIAMI + CLOCK) $\to$ CLOCK	0.77	0.88	0.83	0.91
(c) AU12
Study 1: CLOCK $\to$ CLOCK	0.78	0.86	0.76	0.93
Study 2: (MIAMI + CLOCK) $\to$ CLOCK	0.80	0.86	0.77	0.93
Study 3: MIAMI $\to$ CLOCK	0.74	0.80	0.70	0.92
Hammal et al. (2017)	0.77	−	0.64	0.93
OpenFace	0.67	0.83	0.69	0.89
Study 4 - Adult AFAR: (EB+ + GFT) $\to$ CLOCK	0.73	0.76	0.65	0.92
Study 5 - Infant AFAR: (EB+ + GFT + MIAMI + CLOCK) $\to$ CLOCK	0.79	0.87	0.78	0.93
(d) AU20
Study 1: CLOCK $\to$ CLOCK	0.72	0.81	0.66	0.91
Study 2: (MIAMI + CLOCK) $\to$ CLOCK	0.72	0.83	0.67	0.91
Study 3: MIAMI $\to$ CLOCK	0.68	0.57	0.24	0.91
Hammal et al. (2017)	0.72	−	0.48	0.92
OpenFace	0.58	0.58	0.31	0.87
Study 4 - Adult AFAR: (EB+ + GFT) $\to$ CLOCK	−	−	−	−
Study 5 - Infant AFAR: (EB+ + GFT + MIAMI + CLOCK) $\to$ CLOCK	−	−	−	−

Left-side of the $\to$ denotes the database(s) used to train the model in the related study. Right-side of the $\to$ denotes the database used to test the models (i.e. CLOCK)

The best results are shown in bold