CFFNet: Cross-scale Feature Fusion Network for Real-Time Semantic Segmentation

Despite deep learning based semantic segmentation methods have achieved significant progress, the inference speed of high-performance segmentation model is harder to meet the demand of various real-time applications. In this paper, we propose an cross-scale feature fusion network (CFFNet) to harvest the compact segmentatiHon model with high accuracy. Specifically, we design a novel lightweight residual block in backbone with increasing block depth strategy instead of inverted residual block with increasing local layer width strategy for better feature representative learning while reducing the computational cost by about 75%. Moreover, we design the cross-scale feature fusion module which contains three path to effectively fuse semantic features with different resolutions while enhancing multi-scale feature representation via cross-edge connections from inputs to last path. Experiments on Cityscapes demonstrate that CFFNet performs agreeably on accuracy and speed. For 2048 x 1024 input image, our model achieves 81.2% and 79.9% mIoU on validation and test sets at 46.5 FPS on a 2080Ti GPU.