Real-Time Multi-task Network for Autonomous Driving

End-to-end Network has become increasingly important in multi-tasking, especially a driving perception system in autonomous driving. This work systematically introduces an end-to-end perception network for multi-tasking and proposes several key optimizations to improve accuracy. First, we propose efficient segmentation head and box/class prediction networks based on weighted bidirectional feature network. Second, we propose automatically customized anchor for each level in the weighted bidirectional feature network. Third, we propose an efficient training loss function. Based on these optimizations, we develope an end-to-end perception network to perform multi-tasking, including traffic object detection, drivable area segmentation and lane detection simultaneously which achieves better accuracy than prior art. In particular, our network design achieves the state-of-the art 77 mAP@.5 on BDD100K Dataset, outperforms lane detection with 0.293 mIOU on 12.83 parameters and 15.6 FLOPs. The network can perform visual perception tasks in real-time and thus is a practical and accurate solution to the multi-tasking problem.