A CNN-based Multi-Task Regressor Network for Three-dimensional Reconstruction From 3D Point Cloud

The field of three-dimensional reconstruction plays a pivotal role across diverse domains such as computer graphics, virtual reality, robotics, archaeology, and medical imaging. This paper presents a novel deep learning framework for three-dimensional object reconstruction utilizing 3D point clouds, specifically focusing on a Multi-Task Regressor within a Convolutional Neural Network (CNN) architecture. Acknowledging the challenges associated with applying convolutions directly to point clouds, we adopt an innovative approach by building upon the PointNet architecture. Our adaptation integrates a Multi-Task Regressor to achieve highly accurate reconstructions of Supershapes from 3D point clouds. Our methodology involves employing PointNet to extract features from the input 3D point cloud. Subsequently, these features are inputted into a Multi-Task Regressor, which simultaneously predicts multiple outputs. In the context of Supershape reconstruction, the Multi-Task Regressor predicts the essential parameters needed to faithfully recreate the intricate shape. Unlike traditional single-output regressors, our approach considers and predicts multiple facets of the reconstruction process concurrently, enhancing the overall efficiency and accuracy of the model. The results of our experiments surpassed expectations, demonstrating notable improvements in precision and predictive cost. This research not only contributes to the advancement of three-dimensional reconstruction techniques but also highlights the efficacy of employing a Multi-Task Regressor for handling diverse regression objectives within the context of complex shape reconstruction from 3D point clouds.