Graph Neural Network Approaches for 3D Object Reconstruction: A Review of Methods, Benchmarks and Future Directions

Abstract

Graph neural networks (GNNs) have emerged as a powerful and versatile framework for 3D object reconstruction, offering a natural way to model irregular geometric relationships inherent in point clouds, meshes, and other non-Euclidean 3D rep representations. This comprehensive survey examines GNN-based approaches across five interconnected research areas: point cloud completion and up-sampling, mesh and building reconstruction, deformable and dynamic object reconstruction, implicit and hybrid representations, and foundational 3D GNN architectures. We provide an integrated analysis of existing methods and their associated trade-offs, leading to several key insights: (1) GNNs offer a unified geometric reasoning framework across a wide range of 3D vision tasks; (2) domain-specific approaches that explicitly exploit structural priors consistently outperform generic learning-based methods; and (3) accuracy efficiency trade-offs can often be effectively addressed through simplified or carefully constrained architectures. Building on these findings, we identify several promising directions for future research, most notably the growing adoption of hybrid GNN transformer architectures, which combine local geometric reasoning with global contextual modelling and represent an emerging design paradigm for next-generation 3D reconstruction systems. We review benchmark datasets and evaluation metrics, present a comprehensive cross-method comparison, and identify key open challenges, including scalability to large dynamic scenes, standardized implicit-explicit integration, and real-time deployment requirements. This survey synthesizes the current state-of the-art GNNs as a central technology for advancing 3D reconstruction pipelines in computer vision and graphics applications, including autonomous systems, robotics, and medical imaging.