NeB-SLAM: Neural Blocks-based Salable RGB-D SLAM for Unknown Scenes

TL;DR

This work proposes NeB-SLAM, a neural block-based scalable RGB-D SLAM for unknown scenes that represents the entire unknown scene as a set of sub-maps and introduces an adaptive map growth strategy to achieve adaptive allocation of neural blocks during camera tracking and gradually cover the whole unknown scene.

Abstract

Neural implicit representations have recently demonstrated considerable potential in the field of visual simultaneous localization and mapping (SLAM). This is due to their inherent advantages, including low storage overhead and representation continuity. However, these methods necessitate the size of the scene as input, which is impractical for unknown scenes. Consequently, we propose NeB-SLAM, a neural block-based scalable RGB-D SLAM for unknown scenes. Specifically, we first propose a divide-and-conquer mapping strategy that represents the entire unknown scene as a set of sub-maps. These sub-maps are a set of neural blocks of fixed size. Then, we introduce an adaptive map growth strategy to achieve adaptive allocation of neural blocks during camera tracking and gradually cover the whole unknown scene. Finally, extensive evaluations on various datasets demonstrate that our method is competitive in both mapping and tracking when targeting unknown environments.

Figures (7)

• Figure 1: The divide-and-conquer mapping process for unknown scenes. NeBs are adaptively allocated with camera tracking to gradually cover the entire unknown scene.
• Figure 2: Overview of NeB-SLAM. The scene is represented using NeBs, each of which has independent local coordinates and feature encoding. For any 3D point, feature encoding and coordinate encoding are uesd to estimate the sdf value and color by two compact MLPs. Tracking process optimizes the camera pose for each frame, and mapping process jointly optimizes the scene representation and the poses of all keyframes.
• Figure 3: NeB allocation. NeBs are adaptively allocated based on the proportion of newly observed scene to the whole scene in the current viewing frustum.
• Figure 4: Partial reconstruction results on Replica datasetstraub2019replica. All baseline methods are based on known scene size. Our method, however, is capable of obtaining complete, accurate, and high-quality reconstruction results without the need for scene size.
• Figure 5: Qualitative comparison on ScanNetdai2017scannetscene0000 with different shading mode. Our methods achieve accurate scene reconstruction without the need for scene size. In all the figures ground truth trajectory are shown in black and the estimated trajectory are shown in red.