Instance Segmentation Based on Superpixel Module and Attention Module

Abstract

instance segmentation has become a hot topic in computer vision, which is a combination of semantic image segmentation and object detection. In this work, we propose a new network MaskSAN for instance segmentation, which has three outputs: box prediction, semantic segmentation, and classification score. In this network, Superpixel-pooling is used to increase the spatial prior information of features and enhance the edge performance of segmentation. In addition, the network also combines the recently popular attention modules to enhance the feature extraction module of the network and the context information in feature extraction. Without reducing the speed, the accuracy of segmentation and detection is improved. The network makes full use of the image’s own information and enhances the constraints. Our proposed network performs experiments on the public coco dataset and has excellent performance in detection and segmentation.