LAGNet: Lightweight Attention Guided Network for Real-time Semantic Segmentation

Abstract

In recent years, semantic segmentation has been used extensively in a variety of scenarios. It is essential for most practical applications that predictions are accurate and efficient. Toward this end, we present a novel lightweight attention guided semantic segmentation network (LAGNet) aiming for a balance between prediction accuracy and running efficiency. As a first step, we developed the Efficient Inter-attention Bottleneck Module (EIRM) in order to obtain local and contextual information at a lower cost of computation. We then present a novel Image Decomposition Attention Mechanism (IDAM) that refines the feature maps at various stage. Furthermore, we present a novel Decoder named Mutual-Attention Guided Decoder to promote the accuracy of the prediction results, which utilizes attention mechanism to recover the detailed information effectively. The results of extensive experiments on Cityscapes and Camvid datasets show that our model achieves 73.3% and 71.4% mIoU along with 110 and 105 frames per second on the Cityscapes and Camvid datasets, respectively.