Reg-PointNet++: A CNN Network Based on PointNet++ Architecture for 3D Reconstruction of 3D Objects Modeled by Supershapes

Abstract

The use of 3D reconstruction in computer vision applications has opened up new avenues for research and development. It has a significant impact on a range of industries, from healthcare to robotics, by improving the performance and abilities of computer vision systems. In this paper we aim to improve 3D reconstruction quality and accuracy. The objective is to develop a model that can learn to extract features, estimate a Supershape parameters and reconstruct 3D directly from input points cloud. In this regard, we present a continuity of our latest works, using a CNN-based Multi-Output and Multi-Task Regressor, for 3D reconstruction from 3D point cloud. We propose another new approach in order to refine our previous methodology and expand our findings. It is about “Reg-PointNet++”, which is mainly based on a PointNet++ architecture adapted for multi-task regression, with the goal of reconstructing a 3D object modeled by Supershapes from 3D point cloud. Given the difficulties encountered in applying convolution to point clouds, our approach is based on the PointNet ++ architecture. It is used to extract features from the 3D point cloud, which are then fed into a Multi-task Regressor for predicting the Supershape parameters needed to reconstruct the shape. The approach has shown promising results in reconstructing 3D objects modeled by Supershapes, demonstrating improved accuracy and robustness to noise and outperforming existing techniques. Visually, the predicted shapes have a high likelihood with the real shapes, as well as a high accuracy rate in a very reasonable number of iterations. Overall, the approach presented in the paper has the potential to significantly improve the accuracy and efficiency of 3D reconstruction, enabling its use in a wider range of applications.