why is the result of the display of other methods in Fig. 5 very poor?

[nachifur]

For the comparison of segmentation (BDD100K), I pay more attention to how other methods (Such as ToDayGAN. The result of ToDayGAN shown in Fig. 3 is good. But the result of the display of ToDayGAN in Fig. 5 is very poor.) are trained? Because other methods seem to show poor results. Are other methods pre-trained on the Cityscapes dataset?