MDSNet: a lightweight network for real-time vision task on the unmanned mobile robot

Abstract

To makes a trade-off between accuracy and inference speed for semantic segmentation, Multi-Scale Depthwise Separation network (MDSNet) is designed to be effective both in terms of accuracy and inference speed. This network extract local information and contextual information jointly and has feature maps with high spatial resolution. Compared with state-of-the-art algorithms, MDSNet achieves 66.57 MIoU on Camvid with only 0.5M parameters and 79.4 FPS inference speed on a single GTX 1070Ti card. Besides, MDS is deployed on the unmanned platform to test performance under different conditions. The results show that the proposed algorithm performs well on real-time applications in the real world.