Instance Segmentation with BoundaryNet

Instance segmentation is one of the key technology in many domains, such as medical image analysis, traffic and critical infrastructures monitoring, understanding of natural scenes. Recent methods for instance segmentation rely on bounding box regression, however the bounding boxes are not a natural representation for many domains. We address the limitations of the bounding boxes with a new approach called BoundaryNet, in which we train a fully convolutional neural network to draw the boundaries around each object of each class. The boundaries allow for an easy bounding box and mask inference while still providing detailed information about the shape of the object. BoundaryNet avoids the restrictions of YOLO such as the number of bounding boxes, while it is more computationally efficient than the R-CNN methods. The conducted experiments with the proposed neural network architecture BoundaryNet on the Common Object in Context (COCO) dataset show promising results in improving the instance segmentation process.