多模型融合投票预标注算法研究
Research on Pre-labelling Algorithm for Multi-model Fusion Voting
吉星 1陈喆 1陈飞扬 1杨文听 1樊桢珍 1许丹1
作者信息
- 1. 陕西重型汽车有限公司,陕西 西安 710000
- 折叠
摘要
针对标注内容烦琐、耗时等问题,提出一种多模型融合投票预标注方法.在预标注过程中,将Cascade_RCNN、RetinaNet、CondLaneNet三个模型的检测结果进行融合,然后将各个模型生成的坐标结果进行提取、判断、匹配、参数平均、排序等处理,得到最终的预标注结果.在公开数据集以及自建数据集上进行多次试验的结果表明,算法能够提高预标注精度,减少标注过程中人工标注工作量,具有较好的效果,验证了该方法的有效性.
Abstract
Aiming at the two problems of cumbersome and time-consuming annotation content,a pre-labelling algorithm for multi-model fusion voting is proposed.In the pre-labelling process,the detection results of the three models of Cascade_RCNN,RetinaNet and CondLaneNet are fused,and then the coordinate results generated by each model are processed by extracting,judging,matching,averaging of parameters,sorting and so on,to obtain the final pre-labelling results.The results of multiple tests on the public datasets and the self-constructed datasets show that the algorithm is able to improve the accuracy of pre-labelling and reduce the manual labelling workload in the process of labelling,which has a better effect and verifies the effectiveness of the method.
关键词
深度学习/目标检测/车道线检测/预标注/模型融合Key words
Deep Learning/target detection/laneline detection/pre-labelling/model fusion引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据