Deep learning of volumetric representation for 3D object recognition

Abstract

Robust 3D object detection and pose estimation is still a big challenging for robot vision. In this paper, we propose a new framework for 3D object detection and pose estimation. Rather than using RGB-D image as the original data, we propose to use volumetric representation with the help of unsupervised deep learning network to extract low dimensional feature from 3D point cloud directly. The volumetric representation can not only eliminate the dense scale sampling for offline model training, but also reduce the distortion by mapping the 3D shape to 2D plane and overcome the dependence on texture information. Depending on the Hough forest, we can achieve multi-object detection and pose estimation simultaneously. In compare with the state-of-the-arts using public datasets, we justify the effectiveness of our proposed method.