Abstract: Existing real-time semantic segmentation models usually pursue lower inference costs by compressing spatial resolution, reducing the number of feature extraction layers, or simplifying the ...