Hyper d Cityscapes . λ2=0.7,L=LsCE+λ2LtCE d 1 1.5 2 2.5 3 3.5 4 mIoU 45.19 45.53 45.93 46.14 46.01 45.96 45.67

The pixels inimages are spatially coherent, and the featuresfD from the penultimate layer are expected tohave13 effective characteristics for clustering. It28 is empirically effective and usually performs well. The distance loss and CE loss used in Eq.(11) rely on reliable pseudo-labels to guarantee a correct supervision42 imposedonthenetwork.