Mask-FgS: Feature Global Networks and Supsampling for Instance Segmentation

Abstract

The instance segmentation based on deep learning can combine the results of semantic segmentation and target detection to solve more powerful segmentation tasks. In this paper, we address the instance segmentation task with a network called Mask-FgS, which is based on Mask R-CNN. In addition, in our model, the output of the network is still with three branches: box prediction, semantic segmentation, and classification score. Our network proposes a new feature extraction module that improves the performance of the feature pyramid and provides significant improvements in both accuracy and speed. Furthermore, Supsampling replaces traditional bilinear upsampling in the network on mask branches. The proposed network performs experiments in the instance segmentation branch of the coco data set, and the performance of detection and segmentation is better than the current instance segmentation networks.