Diskmask: Focusing Object Features for Accurate Instance Segmentation of Elongated or Overlapping Objects

Abstract

Deep learning has enabled automated segmentation in a large variety of cases. Instance segmentation of touching and overlapping objects remains an open challenge. We present an end-to-end approach that focuses object detections and features to local regions in an encoder stage and derives accurate instance masks in a decoder. We avoid heavy pre- or postprocessing, such as lifting or non-maximum suppression. The approach compares favorably to the current state-of-the-art on three challenging biological datasets.