Real-Time Semantic Segmentation With Fast Attention

In deep CNN based models for semantic segmentation, high accuracy relies on rich spatial context (large receptive fields) and fine details (high resolution), both of which incur computational costs. this letter, we propose a novel architecture that addresses challenges achieves state-of-the-art performance segmentation high-resolution images videos in real-time. The proposed our fast attention, is simple yet efficient modification the popular self-attention mechanism captures same at small fraction cost, by changing order operations. Moreover, to efficiently process input, apply an additional reduction intermediate feature stages network with minimal loss thanks use attention module fuse features. We validate method series experiments, show results multiple datasets demonstrate superior better speed compared existing approaches real-time segmentation. On Cityscapes, 74.4% mIoU 72 FPS 75.5% 58 single Titan X GPU, ~50% faster than while retaining accuracy.