Self-supervised learning approaches for unsupervised domain adaptation (UDA)\nof semantic segmentation models suffer from challenges of predicting and\nselecting reasonable good quality pseudo labels. In this paper, we propose a\nnovel approach of exploiting scale-invariance property of the semantic\nsegmentation model for self-supervised domain adaptation. Our algorithm is\nbased on a reasonable assumption that, in general, regardless of the size of\nthe object and stuff (given context) the semantic labeling should be unchanged.\nWe show that this constraint is violated over the images of the target domain,\nand hence could be used to transfer labels in-between differently scaled\npatches. Specifically, we show that semantic segmentation model produces output\nwith high entropy when presented with scaled-up patches of target domain, in\ncomparison to when presented original size images. These scale-invariant\nexamples are extracted from the most confident images of the target domain.\nDynamic class specific entropy thresholding mechanism is presented to filter\nout unreliable pseudo-labels. Furthermore, we also incorporate the focal loss\nto tackle the problem of class imbalance in self-supervised learning. Extensive\nexperiments have been performed, and results indicate that exploiting the\nscale-invariant labeling, we outperform existing self-supervised based\nstate-of-the-art domain adaptation methods. Specifically, we achieve 1.3% and\n3.8% of lead for GTA5 to Cityscapes and SYNTHIA to Cityscapes with VGG16-FCN8\nbaseline network.\n