Attention-Based Image-to-Video Translation for Synthesizing Facial Expression Using GAN

Table 2

Average content distance, ACD (ACD-I, ACD-C, and ACD-G) scores result of the AffineGAN, MAEGAN, MAEGAN + SA, and the ground truth.

Numbers	Experiment	ACD-I	ACD-C	ACD-G
1	AffineGAN	1.606 ± 0.018	1.452 ± 0.008	1.769 ± 0.007
2	MAEGAN	1.58 ± 0.016	1.414 ± 0.007	1.738 ± 0.011
3	MAEGAN + SA	1.584 ± 0.001	1.430 ± 0.009	1.744 ± 0.01
4	Ground truth	1.38 ± 0.02	1.400 ± 0.01	0

All the ACD scores were also computed for the real (ground truth) for reference, and the last row indicates the result of the ACD scores for the ground truth. The ACD scores are lower, the better the model performance is.