MTNAS: Search Multi-Task Networks for Autonomous Driving

Hao Liu, Dong Li, JinZhang Peng, Qingjie Zhao, Lu Tian, Yi Shan; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

Multi-task learning (MTL) aims to learn shared representations from multiple tasks simultaneously, which has yielded outstanding performance in widespread applications of computer vision. However, existing multi-task approaches often demand manual design on network architectures, including shared backbone and individual branches. In this work, we propose MTNAS, a practical and principled neural architecture search algorithm for multi-task learning. We focus on searching for the overall optimized network architecture with task-specific branches and task-shared backbone. Specifically, the MTNAS pipeline consists of two searching stages: branch search and backbone search. For branch search, we separately optimize each branch structure for each target task. For backbone search, we first design a pre-searching procedure t1o pre-optimize the backbone structure on ImageNet. We observe that searching on such auxiliary large-scale data can not only help learn low-/mid-level features but also offer good initialization of backbone structure. After backbone pre-searching, we further optimize the backbone structure for learning task-shared knowledge under the overall multi-task guidance. We apply MTNAS to joint learning of object detection and semantic segmentation for autonomous driving. Extensive experimental results demonstrate that our searched multi-task model achieves superior performance for each task and consumes less computation complexity compared to prior hand-crafted MTL baselines. Code and searched models will be released at https://github.com/RalphLiu/MTNAS.