EFRNet: Efficient Feature Reuse Network for Real-time Semantic Segmentation

Semantic segmentation is a kind of dense prediction task, which has high requirements on the prediction accuracy and inference speed in mobile terminals. To reduce the computational burden of the segmentation network and supplement the missing spatial information of high-level features, an efficient feature reuse network (EFRNet) is proposed in two steps: a Multi-scale Bottleneck module is designed to extract multi-scale features, and a lightweight backbone is designed based on the MB module; then, features of different depths are integrated through efficient feature reuse model. Experiments on Cityscapes datasets demonstrate that the proposed EFRNet achieves an impressive balance between speed and precision. Specifically, without any pre-trained model and post-processing, it achieves 75.58% Mean IoU on the Cityscapes test dataset with the speed of 118 FPS on a single RTX 2080Ti GPU.