We address the task of active learning in the context of semantic\nsegmentation and show that self-consistency can be a powerful source of\nself-supervision to greatly improve the performance of a data-driven model with\naccess to only a small amount of labeled data. Self-consistency uses the simple\nobservation that the results of semantic segmentation for a specific image\nshould not change under transformations like horizontal flipping (i.e., the\nresults should only be flipped). In other words, the output of a model should\nbe consistent under equivariant transformations. The self-supervisory signal of\nself-consistency is particularly helpful during active learning since the model\nis prone to overfitting when there is only a small amount of labeled training\ndata. In our proposed active learning framework, we iteratively extract small\nimage patches that need to be labeled, by selecting image patches that have\nhigh uncertainty (high entropy) under equivariant transformations. We enforce\npixel-wise self-consistency between the outputs of segmentation network for\neach image and its transformation (horizontally flipped) to utilize the rich\nself-supervisory information and reduce the uncertainty of the network. In this\nway, we are able to find the image patches over which the current model\nstruggles the most to classify. By iteratively training over these difficult\nimage patches, our experiments show that our active learning approach reaches\n$\\sim96\\%$ of the top performance of a model trained on all data, by using only\n$12\\%$ of the total data on benchmark semantic segmentation datasets (e.g.,\nCamVid and Cityscapes).\n