Active learning for semantic segmentation with expected change

摘要: We address the problem of semantic segmentation: classifying each pixel in an image according to class it belongs (e.g. dog, road, car). Most existing methods train from fully supervised images, where is annotated by a label. To reduce annotation effort, recently few weakly approaches emerged. These require only labels indicating which classes are present. Although their performance reaches satisfactory level, there still substantial gap between accuracy and methods. this with novel active learning method specifically suited for setting. model as pairwise CRF cast finding its most informative nodes. nodes induce largest expected change overall state, after revealing true Our criterion equivalent maximizing upper-bound on gain. Experiments two data-sets show that our achieves 97% percent corresponding model, while querying less than 17% (super-)pixel labels.