Abstract
Instance segmentation of biological images is essential for studying objectbehaviors and properties. The challenges, such as clustering, occlusion, andadhesion problems of the objects, make instance segmentation a non-trivialtask. Current box-free instance segmentation methods typically rely on localpixel-level information. Due to a lack of global object view, these methods areprone to over- or under-segmentation. On the contrary, the box-based instancesegmentation methods incorporate object detection into the segmentation,performing better in identifying the individual instances. In this paper, wepropose a new box-based instance segmentation method. Mainly, we locate theobject bounding boxes from their center points. The object features aresubsequently reused in the segmentation branch as a guide to separate theclustered instances within an RoI patch. Along with the instance normalization,the model is able to recover the target object distribution and suppress thedistribution of neighboring attached objects. Consequently, the proposed modelperforms excellently in segmenting the clustered objects while retaining thetarget object details. The proposed method achieves state-of-the-artperformances on three biological datasets: cell nuclei, plant phenotypingdataset, and neural cells.