Introduction
Autonomous robots require accurate environmental models for effective localization and planning. While traditional explicit representations like voxel grids or point clouds are common, they often suffer from a trade-off between memory consumption and detail, especially in large-scale scenes.
Neural implicit representations, which store geometry as a continuous function (such as a Signed Distance Function or SDF), have emerged as a compact alternative. However, existing implicit methods still face challenges with memory bloat in large environments and artifacts caused by sparse or noisy LiDAR data. 3QFP addresses these issues by introducing a representation that is both extremely sparse and capable of smooth interpolation.
Methods
3QFP introduces two primary innovations to optimize neural implicit mapping:
- Tri-Quadtrees: Rather than utilizing dense 3D voxel grids or dense 2D planes, the system projects the scene onto three orthogonal planes (XY, XZ, YZ). These planes are represented using sparse quadtrees, combining the memory efficiency of hierarchical structures with the computational simplicity of planar projections. Features are only stored where surfaces exist, drastically reducing redundancy.
- Hybrid Feature Encoding: To query a point, features are bilinearly interpolated from the quadtrees and concatenated with a Fourier feature positional encoding. This combined vector is processed by a tiny Multi-Layer Perceptron (MLP) to predict the SDF value. The Fourier encoding provides a global inductive bias that helps the model "fill in the blanks" in regions with sparse data.
The model is trained via test-time optimization. As LiDAR scans are received, quadtree features and MLP weights are updated end-to-end using a Binary Cross Entropy (BCE) loss to distinguish between occupied surfaces and free space.
Experimental Evaluation
The authors evaluated 3QFP across three distinct datasets:
- MaiCity: A synthetic urban dataset providing clean ground truth for quantitative analysis.
- NewerCollege: A real-world dataset from a college campus used to test performance in noisy, complex environments.
- KITTI (Seq07): Used for large-scale qualitative visual comparisons.
Key Results
3QFP demonstrates significant improvements in both efficiency and robustness:
- Memory Efficiency: The method requires only 10%–20% of the memory used by leading neural implicit baselines like SHINE-Mapping. For example, on the NewerCollege dataset, 3QFP utilized approximately 1.6M parameters compared to 11.4M for SHINE-Mapping.
- Robustness to Sparsity: In experiments where input data was subsampled (e.g., using every 6th frame), 3QFP maintained a high completion ratio. In contrast, explicit methods like VDBFusion experienced a sharp decline in map quality under the same conditions.
- Surface Quality: The combination of learnable features and Fourier encoding leads to significantly smoother surfaces and fewer holes compared to methods relying solely on local features.
Conclusion and Future Work
3QFP provides a highly efficient framework for large-scale surface reconstruction. By leveraging tri-quadtree sparsity and Fourier encodings, it achieves state-of-the-art results with a fraction of the memory footprint.
Future work includes integrating this representation into a full SLAM (Simultaneous Localization and Mapping) pipeline to handle unknown robot poses, investigating multi-MLP architectures for even larger scales, and extending the framework to handle dynamic objects in moving scenes.
Summary updated