A Weakly Supervised Amodal Segmenter with Boundary Uncertainty Estimation

This paper addresses weakly supervised amodal instance segmentation, where\nthe goal is to segment both visible and occluded (amodal) object parts, while\ntraining provides only ground-truth visible (modal) segmentations. Following\nprior work, we use data manipulation to generate occlusions in training images\nand thus train a segmenter to predict amodal segmentations of the manipulated\ndata. The resulting predictions on training images are taken as the\npseudo-ground truth for the standard training of Mask-RCNN, which we use for\namodal instance segmentation of test images. For generating the pseudo-ground\ntruth, we specify a new Amodal Segmenter based on Boundary Uncertainty\nestimation (ASBU) and make two contributions. First, while prior work uses the\noccluder's mask, our ASBU uses the occlusion boundary as input. Second, ASBU\nestimates an uncertainty map of the prediction. The estimated uncertainty\nregularizes learning such that lower segmentation loss is incurred on regions\nwith high uncertainty. ASBU achieves significant performance improvement\nrelative to the state of the art on the COCOA and KINS datasets in three tasks:\namodal instance segmentation, amodal completion, and ordering recovery.\n