D-GCNN: Disentangled Graph Convolutional Neural Network Point Cloud Ghosting Removal

• Main Authors: GOU-WANG, DING-ZHI, 郭王鼎志
• Other Authors: HUANG, CHING-CHUN
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/txe2bw

碩士 === 國立中正大學 === 電機工程研究所 === 107 === In order to achieve the 3D information, many cheap depth sensors can be used such as Microsoft's Kinect and Intel's RealSence. However, with these cheap sensors, many measurement errors may be produced, which may cause ghosting problem in the 3D reconstruction results when matching the 3D point clouds from different viewing angles. In this work, to handle this problem, we propose a novel Disentangled Graph Convolutional Network (D-GCN) for point cloud ghosting removal. In this network, GCN is first applied to capture the local structural information of the 3D reconstruction result. Then, these information will be decomposed into two parts: object information, and ghosting information by disentangled representation learning. With the proposed network, the ghosting effect later on can be removed by eliminating the ghosting information before doing reconstruction. Experimental results on synthetic datasets show that the proposed D-GCN significantly outperforms the traditional point cloud-based 3D reconstruction methods. The proposed network structure also represents the potential to become a generic architecture for reconstructing a complete point cloud from an arbitrary ghosting point cloud. Keywords: 3D reconstruction, Ghosting removal, Graph convolutional networks, Disentangled representation learning.