Han Cheng ; Zhao Qichao ; Zhang Shuyi ; Chen Yinzi ; Zhang Zhenlin ; Yuan Jinwei

Abstract

Over the last decade, multi-tasking learning approaches have achievedpromising results in solving panoptic driving perception problems, providingboth high-precision and high-efficiency performance. It has become a popularparadigm when designing networks for real-time practical autonomous drivingsystem, where computation resources are limited. This paper proposed aneffective and efficient multi-task learning network to simultaneously performthe task of traffic object detection, drivable road area segmentation and lanedetection. Our model achieved the new state-of-the-art (SOTA) performance interms of accuracy and speed on the challenging BDD100K dataset. Especially, theinference time is reduced by half compared to the previous SOTA model. Codewill be released in the near future.