四维标注在智能驾驶BEV感知中的应用

Application of 4D labelling in intelligent driving BEV perception

感知是智能驾驶应用中的重要一环,随着BEV感知算法的推进,对数据标注的质量和数量级提出更高的要求。本文深入研究BEV感知算法对数据标注的详细需求,并结合当前主流数据标注业务形态,提出一种基于向量空间的四维数据标注方法,将标注结果实时投影至图像空间,真正做到标注一次,输出海量样本。实践结果表明,四维标注不仅可以提升标注效率百倍以上,还能从本质上保障数据标注质量,应对BEV感知需求。

Perception is an important part of intelligent driving application.With the advancement of BEV(bird s-eye view)perception algorithm in the industry,higher requirements are put forward for the quality and order of magnitude of data annotation.In this paper,the detailed requirements of BEV perception algorithm for data annotation are deeply studied,and a new 4D data annotation method upon vector space is proposed based on the weakness of current mainstream data annotation methods.The annotation results are projected into the image space in real time,and the output of massive training samples can be truly annotated once.Practical results show that 4D annotation can not only improve the annotation efficiency more than a hundred times,but also guarantee the quality of data annotation in essence and meet the needs of BEV perception.