基于SfM单目六自由度位姿估计数据集自动标注
Annotate monocular 6DoF pose estimation datasets based on SfM
刘毅 1魏东辰 2李子豪 1严小军3
作者信息
- 1. 北京航天控制仪器研究所工艺技术研究室,北京 100854
- 2. 北京易航远智科技有限公司感知部,北京 100015
- 3. 北京航天控制仪器研究所 测试技术研究室,北京 100854
- 折叠
摘要
为解决在训练物体六自由度位姿估计神经网络时,人工标注真实场景数据集困难的问题,提出一种自动生成大量单目六自由度位姿估计数据集的方法,可提高数据集标注效率和精度.考虑采集图象环境的光照、物体遮挡等条件,以单目RGB相机、物体三维模型作为输入,在运动恢复结构(structure form motion,SfM)算法框架中添加尺度先验信息约束,实现在真实场景快速生成大量用于六自由度位姿估计训练的数据集.以生活用品为例,分别制作无遮挡、有遮挡数据集,与现有六自由度位姿估计数据集作对比,使用神经网络算法验证根据该方法制作出数据集的可行性与有效性.
Abstract
To solve the problem that it is difficult to manually annotate the real scene datasets for 6DoF object pose estimation neural network,an automatic and quick method was suggested to generate a large number of monocular 6D pose estimation data-sets,to improve the efficiency and accuracy of dataset annotation.Based on the structure from motion framework constrained by scale prior information,a large number of datasets in real scenes was rapidly generated from monocular RGB camera and 3D model of objects in the environment with occlusion and illumination.An example was created,and it was used to train a neural network to verify the feasibility and effectiveness of this method.
关键词
数据集/位姿估计/真实场景/深度学习/单目相机/尺度约束/运动恢复结构Key words
datasets/pose estimation/real scene/deep learning/monocular/scale constrained/structure from motion引用本文复制引用
基金项目
国防技术科研基金项目(JCKY2018203A002)
出版年
2024
NETL
NSTL
万方数据