CEREALS - Cost-Effective REgion-based Active Learning for Semantic\n Segmentation

Abstract

State of the art methods for semantic image segmentation are trained in a\nsupervised fashion using a large corpus of fully labeled training images.\nHowever, gathering such a corpus is expensive, due to human annotation effort,\nin contrast to gathering unlabeled data. We propose an active learning-based\nstrategy, called CEREALS, in which a human only has to hand-label a few,\nautomatically selected, regions within an unlabeled image corpus. This\nminimizes human annotation effort while maximizing the performance of a\nsemantic image segmentation method. The automatic selection procedure is\nachieved by: a) using a suitable information measure combined with an estimate\nabout human annotation effort, which is inferred from a learned cost model, and\nb) exploiting the spatial coherency of an image. The performance of CEREALS is\ndemonstrated on Cityscapes, where we are able to reduce the annotation effort\nto 17%, while keeping 95% of the mean Intersection over Union (mIoU) of a model\nthat was trained with the fully annotated training set of Cityscapes.\n