Existing neural field-based SLAM methods typically employ a single monolithic field as their scene representation. This prevents efficient incorporation of loop closure constraints and limits scalability. To address these shortcomings, we propose a neural mapping framework which anchors lightweight neural fields to the pose graph of a sparse visual SLAM system. Our approach shows the ability to integrate large-scale loop closures, while limiting necessary reintegration. Furthermore, we verify the scalability of our approach by demonstrating successful building-scale mapping taking multiple loop closures into account during the optimization, and show that our method outperforms existing state-of-the-art approaches on large scenes in terms of quality and runtime.
Method
Represent scene by a set of neural fields anchored to pose graph of sparse visual SLAM system
Each field captures scene within a fixed radius surrounding it
When pose graph updates, the neural fields' poses are updated with it
For efficiency, each field is trained independently
For queries, overlapping areas can be averaged to avoid transition artifacts
