Neural Graph Mapping for Scalable and Robust Dense SLAM with Efficient Loop Closure Integration

The authors propose a novel SLAM framework that combines the robust tracking and efficient loop closure handling of sparse visual SLAM with the differentiable-rendering-based dense mapping of neural scene representations. Their approach represents the scene using an extendable set of lightweight neural fields anchored to the keyframes of the SLAM system's pose graph, enabling large-scale map deformations while limiting necessary reintegration.