SGSPose: Neuromorphic-Geometric 6D Pose Estimation Through Spiking Graph Neural Networks and SE(3)-Equivariant Learning

Accurate estimation of 6D poses, including both the spatial position and orientation of objects or cameras, is a significant challenge in robotics, augmented reality, and autonomous systems. This study introduces SGSPose, a novel architecture that integrates spiking neural network (SNN) feature encoding, graph neural network (GNN) relational reasoning, and SE(3)-equivariant Lie algebra optimization. By leveraging both the event-driven efficiency of SNNs and geometric consistency of SE(3)-equivariant GNNs, SGSPose achieves robust and energy-efficient pose estimation in complex real-world environments. Extensive evaluation of the challenging 7Scenes dataset demonstrates that the SGSPose method achieves state-of-the-art translation accuracy on the 7Scenes benchmark, reducing mean translation errors by over 70% compared to existing techniques and attaining mean rotation errors within 8–21°, which is within 10-30% of the leading methods across scenes, demonstrating its robustness and effectiveness for precise 6-DoF camera relocalization in complex indoor environments. These results highlight the potential of brain-inspired geometry-aware deep learning frameworks for advancing the next generation of 6D pose estimation technologies.