Abstract

Existing methods for pixel-wise labelling tasks generallydisregard the underlying structure of labellings, often lead-ing to predictions that are visually implausible. While in-corporating structure into the model should improve pre-diction quality, doing so is challenging – manually speci-fying the form of structural constraints may be impracticaland inference often becomes intractable even if structuralconstraints are given. We sidestep this problem by reducingstructured prediction to a sequence of unconstrained pre-diction problems and demonstrate that this approach is ca-pable of automatically discovering priors on shape, conti-guity of region predictions and smoothness of region con-tours from data without any a priori specification. On theinstance segmentation task, this method outperforms thestate-of-the-art, achieving a mean APr of 63.6% at 50% overlap and 43.3% at 70% overlap.