Fourier-Based Equivariant Graph Neural Networks for Camera Pose Estimation

Traditional graph neural networks often face challenges in accurately modeling nonlinear transformations, such as rotation and translation, commonly encountered in real-world applications like robot navigation and augmented reality. Our approach enhances pose estimation accuracy by implementing equivariant encoding of nodes and edges within the graph structure, ensuring robust feature representation under various geometric transformations. Additionally, we use an equivariant spatial attention mechanism to facilitate the interaction of node information within the graph, improving accuracy in complex scenarios. Experimental results show our method surpasses the existing state-of-the-art.