Weakly Supervised Instance Segmentation by Learning Annotation Consistent Instances

Recent approaches for weakly supervised instance segmentations depend on two\ncomponents: (i) a pseudo label generation model that provides instances which\nare consistent with a given annotation; and (ii) an instance segmentation\nmodel, which is trained in a supervised manner using the pseudo labels as\nground-truth. Unlike previous approaches, we explicitly model the uncertainty\nin the pseudo label generation process using a conditional distribution. The\nsamples drawn from our conditional distribution provide accurate pseudo labels\ndue to the use of semantic class aware unary terms, boundary aware pairwise\nsmoothness terms, and annotation aware higher order terms. Furthermore, we\nrepresent the instance segmentation model as an annotation agnostic prediction\ndistribution. In contrast to previous methods, our representation allows us to\ndefine a joint probabilistic learning objective that minimizes the\ndissimilarity between the two distributions. Our approach achieves state of the\nart results on the PASCAL VOC 2012 data set, outperforming the best baseline by\n4.2% mAP@0.5 and 4.8% mAP@0.75.\n