Method for augmenting 3D point cloud models using graph neural networks
Ivan Chukhran, Serhii Udovenko, Vadim Shergin, Larysa Chala · Innovative technologies and scientific solutions for industries · 2025
Constructing models representing three-dimensional objects as a set of unstructured points (3D point cloud models) in space is becoming increasingly widespread across various domains, including autonomous navigation, robotics, virtual/augmented reality, and 3D reconstruction. Accurately capturing and processing 3D point cloud data is critical for applications requiring a comprehensive understanding of the surrounding environment, such as obstacle avoidance, path planning, and scene modelling. However, due to various reasons, point clouds often contain missing regions, posing significant challenges for subsequent data processing. Incomplete point cloud data can have serious consequences, for instance, in autonomous navigation systems, where errors may lead to collisions or other hazardous situations. Addressing this issue is crucial for the reliable processing of 3D data. This work aims to develop and investigate a method for automatically completing and reconstructing point clouds using graph neural networks. The study's primary objectives include analysing existing approaches to constructing and restoring three-dimensional graph models, developing and implementing a method for automatically completing point clouds using graph neural networks and modelling the proposed method for tasks related to the completion and 3D reconstruction of point cloud models. In this work, a conceptual model for point cloud completion was developed using graph neural networks, enabling the efficient encoding of incomplete point clouds as graphs and the prediction of missing points. The proposed solution for completing incomplete 3D point clouds offers scientific novelty and combines the power of graph neural networks (GNN) with the Point Completion Network (PCN) architecture. The suggested approach allows for high-quality restoration of incomplete 3D data, essential for numerous applications, such as 3D reconstruction, robot navigation, and more. The practical significance of the work’s results is validated by the modelling outcomes of the developed method on classical datasets and their comparison with existing approaches to solving the studied problem. A promising direction for further research on this topic includes testing various architectures of graph neural networks, tuning hyperparameters, applying alternative loss functions, and leveraging more powerful computational resources to train the constructed neural network models.