Proposal-Free Volumetric Instance Segmentation from Latent Single-Instance Masks

This work introduces a new proposal-free instance segmentation method that\nbuilds on single-instance segmentation masks predicted across the entire image\nin a sliding window style. In contrast to related approaches, our method\nconcurrently predicts all masks, one for each pixel, and thus resolves any\nconflict jointly across the entire image. Specifically, predictions from\noverlapping masks are combined into edge weights of a signed graph that is\nsubsequently partitioned to obtain all final instances concurrently. The result\nis a parameter-free method that is strongly robust to noise and prioritizes\npredictions with the highest consensus across overlapping masks. All masks are\ndecoded from a low dimensional latent representation, which results in great\nmemory savings strictly required for applications to large volumetric images.\nWe test our method on the challenging CREMI 2016 neuron segmentation benchmark\nwhere it achieves competitive scores.\n