Introduction
In the fields of robotics and autonomous navigation, 3D maps are typically represented as Signed Distance Functions (SDFs) to facilitate localization and path planning. However, high-resolution voxel-based SDFs are notoriously memory-intensive. For instance, a $512^3$ grid can consume hundreds of megabytes, posing a significant challenge for edge devices with limited resources. While compression techniques like Tensor Train (TT) decomposition can mitigate storage issues, traditional workflows require decompressing the data before use, which reintroduces memory bottlenecks at runtime.
TT-SDF2PC addresses this by performing point cloud registration directly within the compressed TT domain. By computing gradients directly on the compressed "cores," the method maintains a small memory footprint throughout the entire localization process.
Methodology: Operating in the Compressed Domain
The core innovation of TT-SDF2PC is the elimination of the decompression step. The environment is stored as a 3D SDF compressed via TT-decomposition, which represents a large $N \times N \times N$ tensor as a sequence of small, low-rank core tensors.
Direct Derivatives and Optimization
Rather than reconstructing the voxel grid to compute finite differences, the authors apply sparse finite-difference operators directly to the small core tensors. This allows for the efficient calculation of partial derivatives ($\nabla_x SDF(x)$) with a computational complexity of $O(NR^2)$, where $R$ is the TT-rank. These derivatives are then utilized in a Gauss-Newton optimization framework to iteratively transform a query point cloud, minimizing the squared SDF values and aligning the points with the map's zero-isolevel surface.
Unlike Truncated SDFs (TSDFs), which only provide gradients near surfaces, this method utilizes the full SDF. This approach provides valid gradients even at a distance from objects, significantly widening the convergence basin and improving robustness to poor initial pose estimates.
Experimental Evaluation
The researchers evaluated TT-SDF2PC using two datasets:
- ModelNet: A synthetic dataset of detailed 3D objects (e.g., chairs) used to validate the algorithm under noise-free conditions.
- 3DMatch: A collection of real-world RGB-D data (including 7-scenes and BundleFusion) characterized by sensor noise and reconstruction artifacts.
Results and Discussion
The experimental results demonstrate that TT-SDF2PC achieves a memory reduction of 10x to 200x compared to standard voxel grids, typically requiring only 1–3 MB for high-resolution scenes. Despite this massive compression, the registration accuracy remains competitive with state-of-the-art local methods like point-to-plane ICP.
Key Insights
- Natural Filtering: On real-world data, the low-rank TT-compression acts as a natural low-pass filter. This smoothing effect removes high-frequency sensor noise, often leading to cleaner gradients and better registration performance than the raw, uncompressed SDF.
- Rank Saturation: Accuracy was found to saturate at a relatively low TT-rank (approximately $R=10$). This suggests that coarse geometric features captured in a highly compressed state are sufficient for reliable localization.
- Robustness: The use of the full SDF allows the system to handle significant initial pose perturbations—up to 1.0m in translation and 0.25 rad in rotation—outperforming TSDF-based alternatives.
Conclusion
TT-SDF2PC represents a significant step toward memory-efficient spatial AI. By enabling direct optimization in the compressed domain, it allows high-resolution 3D maps to be used on resource-constrained hardware without sacrificing accuracy or robustness. Future work will focus on closing the loop by integrating dynamic map updates and real-time compression into the pipeline.
Summary updated