Reproduce “TESGNN: Temporal Equivariant Scene Graph Neural Networks for Efficient and Robust Multi-View 3D Scene Understanding” on OpenResearch

About this paper

Quang P. M. Pham, Khoi T. N. Nguyen, Lan C. Ngo, Truong Do, Dezhen Song, Truong-Son Hy

TESGNN, developed by researchers from Mohamed bin Zayed University of Artificial Intelligence and VinUniversity, introduces a Temporal Equivariant Scene Graph Neural Network that leverages symmetry-preserving GNNs and a novel temporal graph matching mechanism to achieve robust and efficient 3D scene understanding from multi-view point clouds. It demonstrates improved accuracy in object and relationship prediction, faster convergence, and enhanced computational efficiency compared to prior state-of-the-art methods.

Key insights

• Explicitly preserving Euclidean symmetry in GNNs significantly enhances the robustness and consistency of 3D scene graph representations, particularly when dealing with noisy, multi-view data.
• Leveraging symmetry-preserving node and edge embeddings, combined with a triplet-based representation and contrastive learning, enables robust and efficient temporal graph matching for global scene understanding.
• A carefully designed GNN architecture that combines attention-based and equivariant layers can improve accuracy and training convergence while reducing computational overhead for 3D scene graph generation.

Summary by alphaXiv.