Stripe Pooling Attention for Real-Time Semantic Segmentation

In order to solve the problem that it is difficult to meet the application in real scene because of the attention mechanism semantic segmentation algorithm cannot achieves a good balance between segmentation speed and accuracy. We proposed a lightweight real-time semantic segmentation algorithm based on strip-pooling attention. Firstly, lightweight backbone network was used to extract feature information, and a feature fusion module was constructed to obtain context information at different scales to improve the segmentation accuracy. Secondly the attention-based strip attention module (SAM) is used to improve the attentiveness of remote information, and horizontal strip pooling is added to SAM to reduce the computation of encoding global context. Experimental results show that the proposed algorithm can achieve high segmentation accuracy and meet the reai-time requirements, mIoU reached 70.6% on Cityscapes test set and the average segmentation speed is 92 frames per second; mIoU reached 66.4 on CamVid test set and the average segmentation speed is 196 frames per second.