3D Orientation Estimation of Industrial Parts from 2D Images using Neural Networks

被引：4

作者：

Langlois, Julien ^{[1
,2
]}

Mouchere, Harold ^{[1
]}

Normand, Nicolas ^{[1
]}

Viard-Gaudin, Christian ^{[1
]}

机构：

[1] Univ Nantes, Lab Sci Numer Nantes, UMR 6004, Nantes, France

[2] Multitude Technol, Laval, France

来源：

PROCEEDINGS OF THE 7TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION APPLICATIONS AND METHODS (ICPRAM 2018) | 2018年

关键词：

Neural Networks; 3D Pose Estimation; 2D Images; Deep Learning; Quaternions; Geodesic Loss; Rendered Data; Data Augmentation;

D O I：

10.5220/0006597604090416

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

In this paper we propose a pose regression method employing a convolutional neural network (CNN) fed with single 2D images to estimate the 3D orientation of a specific industrial part. The network training dataset is generated by rendering pose-views from a textured CAD model to compensate for the lack of real images and their associated position label. Using several lighting conditions and material reflectances increases the robustness of the prediction and allows to anticipate challenging industrial situations. We show that using a geodesic loss function, the network is able to estimate a rendered view pose with a 5 degrees accuracy while inferring from real images gives visually convincing results suitable for any pose refinement processes.

引用

页码：409 / 416

页数：8