Semisupervised Semantic Segmentation with Mutual Correction Learning

Table 2

Comparison with other methods on Cityscapes under different partition protocols. The segmentation network is DeepLabv3+, and the backbone is ResNet-50. SupOnly represents supervised training, using only labeled data.

Method	ResNet-50
	1/16 (186)	1/8 (372)	1/4 (744)	1/2 (1488)
CCT [13]	66.35	72.46	75.68	76.78
GCT [35]	65.81	71.33	75.30	77.09
CAC [34]	—	69.70	72.70	—
CPS [15]	74.47	76.61	77.83	78.77
Ours	74.47	76.75	78.03	78.89

The meaning of the bold values represent the best results.