Efficient fully convolutional network for real-time semantic segmentation

Considering the increased interest for deploying semantic segmentation models in real-world applications, including autonomous driving, robotics, video surveillance, researchers are developing ultra-lightweight segmentation frameworks, comprising approximately 1 million trainable parameters. Therefore, in this study, we propose an efficient convolutional segmentation network, termed as ECSNet, which achieves a good trade-off between segmentation accuracy, inference speed, computational cost, and parameter count. Particularly, we design the dilated convolutional feature module (DCFM), that is integrated within the encoder structure, and it effectively captures the local details and the global information. The decoder design is lightweight, comprising group convolutional layers, concatenation operation, upsampling operation, and the segmentation head, without any additional context extractor module. Comprehensive experiments are conducted on two autonomous driving datasets, namely Cityscapes and CamVid, demonstrating that the proposed framework attains an effective accuracy-speed trade-off. With only 1.06M parameters, ECSNet achieves $$70.98\%$$ mIoU and $$73.45\%$$ mIoU at an inference speed of 294.22 FPS and 234.48 FPS on the Cityscapes and CamVid test sets, respectively.