Vectorized instance segmentation using periodic B-splines based on cascade architecture

Abstract

Instance segmentation is an essential part of image semantic understanding. In this paper we propose a novel cascade framework for instance segmentation. Unlike some existing methods that only output contour discrete coordinates, our approach obtains a vectorized representation of contours using periodic B-splines. In order to make better use of geometry and appearance information, we consider the global and local features of objects and introduce two types of graph structures, the star graph and circular graph, for feature extraction. Thereby, we develop a neural network, termed the mix network, to better exploit extracted features. Specifically, we first regress the spline control points to an object boundary via the mix network, then perform spline sampling to obtain the initial predictions of contours, and finally deform the predicted contours to the real contours of the objects. In addition, we add a regularization to further constrain the fairness of contour splines. Experiments show that our approach achieves 34.6% in mask mAP, Mean Average Precision, with a ResNet-101-FPN-DCN backbone on the challenging COCO benchmark, which significantly improves the performance of contour-based methods.