Systems and methods for performing instance segmentation

Abstract

Systems and methods for performing instance segmentation. A memory stores instructions for executing processes for performing instance segmentation and a processor configured to execute the instructions. The processes include: generating a learning objective that uses pair-wise relationships between pixels in an input image; sampling pixels in each object instance to determine whether the sampled pixels are within a same object instance; training a neural network using the learning objection, wherein the neural network is configured to make pixel-wise predictions and to assign a cluster index to each pixel of the input image, with each pixel cluster being an object instance; performing graph coloring to assign a color to each object instance, with adjacent object instances having different colors; performing connected component extraction to recover each object instance based on the graph coloring; and generating a rendered image having the assigned color applied to each object instance.