Research on Three-Dimensional Point Cloud Reconstruction Method Based on Graph Neural Networks

The widespread utilization of three-dimensional reconstruction technology across various domains such as medicine, architecture, and transportation has led to a growing demand for precise positioning and high-precision modeling of targets. Three-dimensional reconstruction methodologies leveraging deep learning techniques have demonstrated significant advantages. Nonetheless, traditional three-dimensional reconstruction networks often encounter challenges related to loss of intricate image features, low point cloud density, and susceptibility to generating voids, consequently impacting quality and accuracy of three-dimensional reconstructions. To combat this challenge, this study introduces an algorithm based on graph neural networks for dynamically selecting central points as a replacement for the original point cloud enhancement strategy. Initially, the algorithm employs CAS method to identify appropriate central points that cover expanded spatial volumes within their respective neighborhoods. Subsequently, PointFlow algorithm is applied to forecast point clouds at these central points. The resulting point clouds are then refined by integrating original point cloud segments with interpolated point cloud segments, culminating in comprehensive, high-density three-dimensional point clouds representing the target scene. Relative to Point-MVSNet, the algorithm presented in this paper demonstrates a substantial 6.5% reduction in the average error of the reconstructed three-dimensional models. Notably, the resulting point cloud density and fidelity are elevated, showcasing richer detailed features while also exhibiting lower resource consumption compared to alternative three-dimensional reconstruction algorithms.