Recurrent Level Set Networks for Instance Segmentation

Level set (LS)-based segmentation has been widely used in medical imaging domain. It however has some difficulty when dealing with multi-instance objects in the real world. Furthermore, LS’s performance is generally quite sensitive to some initial settings and parameters such as the number of iterations. To address these issues and promote the classic LS methods to a new degree of performance in a trainable deep learning framework, we are presenting a novel approach contextual recurrent level sets (CRLS) for object instance segmentation. In the proposed networks, the curve deformation process is formed as a hidden state evolution procedure in gated recurrent units (GRUs) and updated by minimizing an energy functional composed of fitting forces and contour length.