Efficient semantic segmentation with pyramidal fusion

Marin Orsic, Sinisa Segvic

Abstract: Highlights•We present pyramidal fusion: a principled approach for dense recognition based on resolution pyramids.•Our pyramidal model outperforms all previous semantic segmentation approaches aiming at real-time operation.•It achieves 76.4% mIoU on Cityscapes test by processing 2MPx images at 34 Hz on GTX 1080 Ti.•Our model acts as an ensemble of shallow models with a large effective receptive field.