ACANet: across-scale context attention network for real-time semantic segmentation

With the recent advance of various context aggregation approaches, remarkable progress has been achieved in semantic segmentation. However, it is still challenging to fully exploit the discriminative across-scale context information in an efficient manner. In this paper, we introduce an across-scale context attention network (ACANet) for real-time semantic segmentation. Instead of compute complex query-dependent attention map, we calculate query-independent attention map to aggregate contexts. Experimental results on Cityscape and Camvid datasets demonstrate the effectiveness of our method. In particular, our network achieves 77.4% on the Cityscape test set with a 32 FPS for 1024×2048 images on a single RTX 2080Ti GPU.