MISO: Multiresolution Submap Optimization for Efficient Globally Consistent Neural Implicit Reconstruction

TL;DR

MISO introduces a hierarchical, submap-based back-end for neural implicit SLAM that operates on multiresolution implicit features. Local SLAM optimizes submap features and poses with a fixed offline-trained decoder, while global alignment/fusion refines submap bases and fuses maps in feature space through level-wise, hierarchical optimization, aided by offline pre-trained encoders for rapid initialization. Across ScanNet, FastCaMo-Large, and Newer College datasets, MISO achieves competitive or superior estimation quality with substantial speed-ups over existing methods and robust drift correction without decoding full geometry. The approach significantly improves scalability and robustness of neural SDF SLAM in large-scale real-world environments, with practical impact for real-time mapping and long-term consistency.

Abstract

Neural implicit representations have had a significant impact on simultaneous localization and mapping (SLAM) by enabling robots to build continuous, differentiable, and high-fidelity 3D maps from sensor data. However, as the scale and complexity of the environment increase, neural SLAM approaches face renewed challenges in the back-end optimization process to keep up with runtime requirements and maintain global consistency. We introduce MISO, a hierarchical optimization approach that leverages multiresolution submaps to achieve efficient and scalable neural implicit reconstruction. For local SLAM within each submap, we develop a hierarchical optimization scheme with learned initialization that substantially reduces the time needed to optimize the implicit submap features. To correct estimation drift globally, we develop a hierarchical method to align and fuse the multiresolution submaps, leading to substantial acceleration by avoiding the need to decode the full scene geometry. MISO significantly improves computational efficiency and estimation accuracy of neural signed distance function (SDF) SLAM on large-scale real-world benchmarks.

Figures (11)

• Figure 1: Demonstration of MISO on the FastCaMo-Large dataset tang2023mips. MISO leverages multiresolution submaps that organize neural implicit features at different spatial resolutions (visualized in (a) and (b) using principal component analysis). By performing hierarchical optimization within (local) and across (global) submaps, MISO can efficiently and accurately reconstruct SDF (c) and estimate mesh and robot trajectory (d). For clarity, we only visualize the features and SDF values within $30$ cm of the surface. The scene size is $26.0$ m $\times$$16.7$ m $\times$$7.5$ m.
• Figure 2: Overview of MISO. (a) Given point cloud observations, MISO performs local hierarchical SLAM within a submap represented as a multiresolution feature grid (\ref{['sec:local_optimization']}). (b) Given locally optimized submaps, MISO performs global alignment and fusion across submaps to eliminate estimation drift and achieve globally consistent scene reconstruction (\ref{['sec:global_optimization']}).
• Figure 3: Illustration of the level-$l$ encoder $E_{\phi_l}$. Input point cloud with residuals $\{x_j, r_j^\text{in}\}_j$ is voxelized via averaging pooling and processed by a 3D CNN. The CNN outputs at all vertices are then transformed via a shared MLP to predict the target feature grid $F_l$.
• Figure 4: Visualization of estimated SDF at a fixed height on ScanNet scene 0207. MISO performs SLAM using noisy poses. iSDF and Neural Points use ground truth poses and only perform mapping.
• Figure 5: Inputs and output predictions from learned hierarchical encoders on ScanNet scene 0024. Red and blue show positive and negative residual or SDF values, respectively.

Theorems & Definitions (4)

• Definition 1: Multiresolution Feature Grid
• Proposition 1: Linear least squares
• proof
• proof : Proof of \ref{['lem:linear_case']}