MetaBox+: A new Region Based Active Learning Method for Semantic Segmentation using Priority Maps

We present a novel region based active learning method for semantic image\nsegmentation, called MetaBox+. For acquisition, we train a meta regression\nmodel to estimate the segment-wise Intersection over Union (IoU) of each\npredicted segment of unlabeled images. This can be understood as an estimation\nof segment-wise prediction quality. Queried regions are supposed to minimize to\ncompeting targets, i.e., low predicted IoU values / segmentation quality and\nlow estimated annotation costs. For estimating the latter we propose a simple\nbut practical method for annotation cost estimation. We compare our method to\nentropy based methods, where we consider the entropy as uncertainty of the\nprediction. The comparison and analysis of the results provide insights into\nannotation costs as well as robustness and variance of the methods. Numerical\nexperiments conducted with two different networks on the Cityscapes dataset\nclearly demonstrate a reduction of annotation effort compared to random\nacquisition. Noteworthily, we achieve 95%of the mean Intersection over Union\n(mIoU), using MetaBox+ compared to when training with the full dataset, with\nonly 10.47% / 32.01% annotation effort for the two networks, respectively.\n