摘要
数据闭环是智能无人系统获得高可靠、高适应环境感知性能的重要基础,通常包含数据采集、数据预处理、数据标注、模型训练、模型评估、模型部署等环节.随着多模态传感器数据的快速增长,数据闭环面临关键数据筛选、无标注数据信息挖掘、大规模数据高效标注等挑战.围绕数据闭环中的关键问题,介绍了基于主动学习的数据筛选、多视角图像三维感知模型预训练、基于大模型的数据半自动标注等方法,并通过公开数据集上的定量和定性实验,验证了方法的有效性.
Abstract
Data closed-loop is an important foundation for intelligent unmanned systems to obtain highly reliable and high-a-daptive environment perception performance,which usually includes data collection,data preprocessing,data annotation,model training,model evaluation,model deployment,etc.With the rapid growth of multimodal sensor data,data closed-loop faces chal-lenges such as key data filtering,label-free data mining,and efficient annotation of large-scale data.Focusing on the key prob-lems in the data closed-loop,this research first introduces the methods of data filtering based on active learning,pre-training of multi-view image 3D perception model,and semi-automatic data annotation based on large model.Then,through quantitative and qualitative experiments on public datasets,it verifies the effectiveness of the methods proposed.
NETL
NSTL
万方数据