Table 1.

Segmentation results for the Nyuv2 test set in comparison to the baseline network and other state-of-the-art networks. The DDNet was trained with the pretrained VGG16 and the epochs were set to 250 for the best performance. DDNet-Weight A outperformed other networks in mean accuracy, and it was also superior to the baseline model in the mean intersection-over-union.

Methods	Global acc (%)	Mean IoU (%)	Mean acc (%)
FCN [3]	61.5	30.5	42.4
D-CNN [39]	60.3	27.8	39.3
AdaShare [46]	61.3	29.6	-
EDNAS [47]	58.1	22.1	-
Baseline [11]	66.0	32.7	43.4
DDN	65.2	32.3	44.2
DDN-Weight α	65.1	32.9	45.7