An Improved Mask R-CNN with Global Context Modeling for Instance Segmentation

Abstract

Mask R-CNN is a recently proposed state-of-the-art algorithm for object instance segmentation. However, it is still a challenge to improve the accuracy of Mask R-CNN. In this paper, a novel Mask R-CNN approach is proposed. Firstly, we combine global context modeling with Mask R-CNN by capturing long-range dependencies. Secondly, we extend the Mask R-CNN approach by fusing a new simplified non-local neural network and Squeeze-and-Excitation block, which helps to achieve long-range dependencies and adaptively recalibrate the channel feature response by modeling the interdependence between channels. Numerical comparisons between our improved method and original Mask R-CNN solvers, on COCO 2017 dataset, are presented to demonstrate that our algorithm improves bounding box AP by 1.3% and mask AP by 1.6%