Deep Multi-Branch Aggregation Network for Real-Time Semantic Segmentation in Street Scenes

Real-time semantic segmentation, which aims to achieve high segmentation accuracy at real-time inference speed, has received substantial attention over the past few years. However, many state-of-the-art methods tend sacrifice some spatial details or contextual information for fast inference, thus leading degradation in quality. In this paper, we propose a novel Deep Multi-branch Aggregation Network (called DMA-Net) based on encoder-decoder structure perform street scenes. Specifically, first adopt ResNet-18 as encoder efficiently generate various levels of feature maps from different stages convolutions. Then, develop (MAN) decoder effectively aggregate and capture multi-scale information. MAN, lattice enhanced residual block is designed enhance representations network by taking advantage structure. Meanwhile, transformation introduced explicitly transform map neighboring branch before aggregation. Moreover, global context used exploit These key components are tightly combined jointly optimized unified network. Extensive experimental results challenging Cityscapes CamVid datasets demonstrate that our proposed DMA-Net respectively obtains 77.0% 73.6% mean Intersection Union (mIoU) speed 46.7 FPS 119.8 only using single NVIDIA GTX 1080Ti GPU. This shows provides good tradeoff between quality