Efficient pyramid context encoding and feature embedding for semantic segmentation

For reality applications of semantic segmentation, inference speed and memory usage are two important factors. To address these challenges, we propose a lightweight feature pyramid encoding network (FPENet) for segmentation with good trade-off between accuracy speed. We use series (FPE) blocks to encode context at multiple scales in the encoder. Each FPE block consists different depthwise dilated convolutions that perform as spatial extract features reduce computational costs. During training, one-shot neural architecture search algorithm is adopted find optimal structure each from large space small cost. After encoder, mutual embedding upsample module introduced decoder, consisting attention blocks. The encoder-decoder mechanism used help aggregate efficiently high-level low-level details. proposed outperforms existing real-time methods fewer parameters improved on Cityscapes CamVid benchmark datasets. Specifically, it achieved 72.3% mean IoU test set only 0.4 M 192.6 FPS an Nvidia Titan V100 GPU, 73.4% 116.2 when running higher resolution images.