Optimal Configuration of Multi-Task Learning for Autonomous Driving

• DOI: 10.3390/s23249729
• Publication Date: 2023-12-11T19:12:51Z
• Authors: Woomin Jun; Minjun Son; Jisang Yoo; Sungjin Lee
• Keywords: Subnet

For autonomous driving, it is imperative to perform various high-computation image recognition tasks with high accuracy, utilizing diverse sensors perceive the surrounding environment. Specifically, cameras are used lane detection, object and segmentation, and, in absence of lidar, extend inferring 3D information through depth estimation, reconstruction, SLAM. However, accurately processing all these operations real-time for driving under constrained hardware conditions practically unfeasible. In this study, considering characteristics performed by given conditions, we investigated MTL (multi-task learning), which enables parallel execution maximize their speed, memory efficiency. Particularly, study analyzes combinations proposes MDO decision optimization) algorithm, consisting three steps, as a means optimization. initial step, MTS set) selected minimize overall latency while meeting minimum accuracy requirements. Subsequently, additional training shared backbone individual subnets conducted enhance predefined MTS. Finally, both each subnet undergo compression maintaining already secured performance. The experimental results indicate that integrated performance critically important configuration optimization MTL, determined ITC (inter-task correlation). algorithm was designed consider construct multi-task sets exhibit ITC. Furthermore, implementation proposed coupled SSL (semi-supervised learning) based training, resulted significant enhancement. This advancement manifested approximately 12% increase detection mAP performance, 15% improvement 27% reduction latency, surpassing previous three-task learning techniques like YOLOP HybridNet.