Trilateral convolutional neural network for 3D shape reconstruction of objects from a single depth view

• Published in: IET image processing Vol. 13; no. 13; pp. 2457 - 2466
• Main Authors: Rivera, Patricio, Valarezo Añazco, Edwin, Choi, Mun-Taek, Kim, Tae-Seong
• Format: Journal Article
• Language: English
• Published: The Institution of Engineering and Technology 14.11.2019
• Subjects: 3D voxel representation; autoencoder CNN; baseline CNN; belief networks; computer vision; convolutional receptive field; dilated convolutions; fully connected convolutional neural network; generative adversarial reconstruction network; image recognition; image reconstruction; image representation; learning (artificial intelligence); medical image processing; neural nets; object recognition; partial object surface; reconstruction performance; Research Article; shape reconstruction; single depth image; single depth view; solid modelling; superior reconstruction results; three‐dimensional convolutional neural network; trilateral convolutional neural network; Tri‐CNN; object recognition; convolutional receptive field; baseline CNN; partial object surface; autoencoder CNN; shape reconstruction; belief networks; learning (artificial intelligence); medical image processing; reconstruction performance; superior reconstruction results; single depth image; three-dimensional convolutional neural network; dilated convolutions; single depth view; 3D voxel representation; image reconstruction; trilateral convolutional neural network; Tri-CNN; generative adversarial reconstruction network; computer vision; image representation; fully connected convolutional neural network; image recognition; neural nets; solid modelling
• ISSN: 1751-9659; 1751-9667
• DOI: 10.1049/iet-ipr.2019.0532

In this study, the authors propose a novel three-dimensional (3D) convolutional neural network for shape reconstruction via a trilateral convolutional neural network (Tri-CNN) from a single depth view. The proposed approach produces a 3D voxel representation of an object, derived from a partial object surface in a single depth image. The proposed Tri-CNN combines three dilated convolutions in 3D to expand the convolutional receptive field more efficiently to learn shape reconstructions. To evaluate the proposed Tri-CNN in terms of reconstruction performance, the publicly available ShapeNet and Big Data for Grasp Planning data sets are utilised. The reconstruction performance was evaluated against four conventional deep learning approaches: namely, fully connected convolutional neural network, baseline CNN, autoencoder CNN, and a generative adversarial reconstruction network. The proposed experimental results show that Tri-CNN produces superior reconstruction results in terms of intersection over union values and Brier scores with significantly less number of model parameters and memory.