Boundary semantic interactive aggregation network for scene segmentation

Abstract

Deep learning-based road scene segmentation has been paid significantly attention for autonomous driving. Most approaches focus on extracting semantic information of target and neglect boundary feature, resulting in difficultly improving holistically semantics modeling. To handle this challenge, we propose boundary semantic interactive aggregation network (BSIA-Net) to segment road scene, which embeds boundary feature into multi-scale semantic features and aggregates interactive information for promoting semantic understanding. The designed BSIA-Net consists of three parts including semantic flow, boundary flow, and aggregation flow. The semantic flow is used to extract multi-scale semantic information and constructs astrous stack pooling (ASP) module to capture rich context information by expanding the receptive field. The boundary flow is specifically designed to refine low-level boundary feature from edge extracting operator (EEO) by multi-scale semantic features. The aggregation flow constructs interactive aggregation module (IAM) to capture long-range dependencies between the inner objects and boundaries by interactively enhancing intra-class consistency. The BSIA-Net successively achieves 81.1 % and 45.41 % in mean intersection over union on Cityscapes and ADE20K datasets. Extensive experiments compared to some other methods demonstrates the effectiveness and advancement of our method.