Generalizable and Accurate 6D Object Pose Estimation Network

Shouxu Fu, Xiaoning Li, Xiangdong Yu, Lu Cao, Xingxing Li

Abstract: 6D object pose estimation is an important task in computer vision, and the task of estimating 6D object pose from a single RGB image is even more challenging. Many methods use deep learning to acquire 2D feature points from images to establish 2D-3D correspondences, and further predict 6D object pose with Perspective-n-Points (PnP) algorithm. However, most of these methods have problems with inaccurate acquisition of feature points, poor generality of the network and difficulty in end-to-end training of the network. In this paper, we design an end-to-end differentiable network for 6D object pose estimation. We propose Random Offset Distraction (ROD) and Full Convolution Asymmetric Feature Extractor (FCAFE) with the Probabilistic Perspective-n-Points (ProPnP) algorithm to improve the accuracy and robustness of 6D object pose estimation. Experiments show that our method achieves a new state-of-the-art result on the LineMOD dataset, with an accuracy of 97.42% in the ADD(-S) metric. Our approach is also very competitive on the Occlusion LineMOD dataset.