Source-linked AI summary
SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations
Xingguang Zhong, Yue Pan, Jens Behley, Cyrill Stachniss
TL;DR
Large-scale outdoor mapping must balance accuracy and completeness with map memory, and conventional approaches can struggle in sparse observations. SHINE-Mapping uses sparse hierarchical octree features, a shared neural decoder, and regularized incremental updates for LiDAR-based SDF mapping. The paper reports more accurate and complete reconstructions with lower map memory than compared methods.
Problem
Large-scale outdoor maps must be accurate and detailed yet compact enough for memory-limited mobile robots.
Source: §I · p. 1Method
SHINE-Mapping learns sparse hierarchical octree features from LiDAR measurements and decodes them into signed distance values with a shared shallow MLP, using feature-update regularization for incremental mapping.
Source: §I · p. 1, A. Implicit Neural Map Representation · p. 3Results
Takeaways & Limitations
The reported results include better accuracy in densely observed regions, greater completeness in sparsely observed regions, and more compact maps than TSDF fusion-based methods.
Source: §I · p. 2
Abstract
from arXiv · showhide
Accurate mapping of large-scale environments is an essential building block of most outdoor autonomous systems. Challenges of traditional mapping methods include the balance between memory consumption and mapping accuracy. This paper addresses the problem of achieving large-scale 3D reconstruction using implicit representations built from 3D LiDAR measurements. We learn and store implicit features through an octree-based, hierarchical structure, which is sparse and extensible. The implicit features can be turned into signed distance values through a shallow neural network. We leverage binary cross entropy loss to optimize the local features with the 3D measurements as supervision. Based on our implicit representation, we design an incremental mapping system with regularization to tackle the issue of forgetting in continual learning. Our experiments show that our 3D reconstructions are more accurate, complete, and memory-efficient than current state-of-the-art 3D mapping methods.
I. INTRODUCTION
Large-scale outdoor mapping must balance map accuracy and detail against limited onboard memory, while existing methods can be incomplete in sparsely observed areas. SHINE-Mapping uses sparse hierarchical implicit features for incremental LiDAR mapping and reports accuracy, completeness, and memory advantages over existing methods.
I. INTRODUCTION
- Outdoor maps must be compact yet accurate enough for environmental detail, but current large-scale methods can struggle to provide both and to reconstruct sparsely observed areas completely.The paper situates this need in mobile robots’ limited onboard memory and the role of 3D maps in localization and navigation.Source: §I · p. 1
- Implicit neural representations have shown compact, high-fidelity reconstruction, but prior implicit mapping work has focused comparatively little on LiDAR and often on small indoor scenes.Fig. 1 shows an incremental reconstruction on KITTI odometry sequence 00.Source: §I · p. 1, Figure 1 · p. 1
- SHINE-Mapping stores incrementally learnable features in a sparse octree and uses a shared shallow MLP to convert local features into signed distance values.It optimizes local features with a binary cross entropy-based loss and queries geometry by interpolating among features.Source: §I · p. 1
- SHINE-Mapping reports higher accuracy in densely observed regions, greater completeness in sparsely observed regions, and more compact maps than TSDF fusion-based methods.It also reports higher accuracy than a volume-rendering-based implicit neural mapping method.Source: §I · p. 2
II. RELATED WORK
Prior mapping work includes explicit geometric representations and neural implicit representations, with different trade-offs in scene scale and memory. SHINE-Mapping addresses dense feature-grid memory costs and forgetting in incremental implicit mapping with sparse features and feature-update regularization.
II. RELATED WORK
- Explicit geometric maps include surfels, meshes, histograms, octree occupancy, and volumetric TSDF integration, and can support localization and navigation.The TSDF approach popularized by Newcombe et al. enabled real-time reconstruction.Source: §II · p. 2
- Neural implicit representations use MLPs to estimate density for volumetric rendering of views or scene geometry, while NeRF [16] helped inspire their use in mapping.The passage describes this line of work as motivated by novel-view synthesis.Source: §II · p. 2
- iMap [31], iSDF [23], and Neural RGBD represent entire indoor scenes with a single MLP, while NICE-SLAM [40] and Go-SURF [37] combine shallow MLPs with local feature grids for larger scenes.The local-grid approaches are described as faster and more accurate for scenes such as multiple rooms, but their dense voxel structures have enormous memory costs.Source: §II · p. 2
- Replay-based incremental methods iMap [31] and iSDF [23] retain historical keyframes, while NICE-SLAM [40] uses local grids and a fixed decoder but can still face more forgetting as grids grow.Replay stores increasing amounts of data and requires resampling mechanisms; SHINE-Mapping instead uses feature-update regularization.Source: §II · p. 2
III. OUR APPROACH – SHINE-MAPPING
SHINE-Mapping represents a continuous signed distance field with learnable hierarchical octree features and a shared MLP decoder. Its sparse, extensible feature storage supports online mapping, with feature optimization and decoding performed across hierarchy levels.
III. OUR APPROACH – SHINE-MAPPING
- The map uses learnable octree-based hierarchical feature grids and a globally shared MLP decoder to represent the environment’s continuous signed distance field.Point clouds with known poses provide direct measurements to supervise the SDF output as local features are optimized online.Source: §III · p. 2
A. Implicit Neural Map Representation
- The octree is built directly from the point cloud, and one hash table per level stores features addressed with unique Morton codes.This permits extension without preallocating memory, maintains the H most-detailed levels, and provides average-case O(1) feature access.Source: A. Implicit Neural Map Representation · p. 3
- Feature vectors are stored at tree-node corners and combined from up to H hierarchical levels by trilinear interpolation and summation.The resulting vector is fed to a shared shallow MLP to obtain the SDF; the differentiable process jointly optimizes feature vectors and MLP parameters.Source: A. Implicit Neural Map Representation · p. 3
- For incremental mapping, the MLP is pretrained and kept fixed to minimize catastrophic forgetting, whereas batch mapping can operate without pretraining.The hierarchy combines features at multiple resolutions, with H equal to 3 or 4 reported as sufficient for good results.Source: A. Implicit Neural Map Representation · p. 3
B. Training Pairs and Loss Function
The method samples along LiDAR rays and uses projected signed distance to supervise SDF predictions. Its sigmoid-mapped binary cross entropy emphasizes samples near the surface, with an additional Eikonal term for batch-based feature optimization.
B. Training Pairs and Loss Function
- Training pairs come from points sampled along LiDAR rays, using the signed distance from each sampled point to the beam endpoint as supervision.The method samples Nf points uniformly in free space and Ns points inside the ±3σ truncation band around the surface.Source: B. Training Pairs and Loss Function · pp. 3–4
- Projected signed distances from sampled points to the surface are mapped through a sigmoid to supervision labels, and binary cross entropy trains the model’s sigmoid-mapped SDF output.The sigmoid’s flatness is controlled by σ, which indicates measurement-noise magnitude; the loss is more sensitive near the surface.Source: B. Training Pairs and Loss Function · p. 3
- Batch-based feature optimization adds an Eikonal term to the loss to encourage accurate signed distance values.The Eikonal weight is λe, and the input-sample gradient is calculated through automatic differentiation.Source: B. Training Pairs and Loss Function · p. 4
C. Incremental Mapping Without Forgetting
Incremental updates can overwrite shared features and reduce the accuracy of previously mapped areas. The method adds importance-weighted regularization to limit changes to features learned from earlier scans.
C. Incremental Mapping Without Forgetting
- Updating shared features to fit a new area can reduce reconstruction accuracy in previously mapped areas, with forgetting becoming more severe as grid size increases.At successive frames, later observations may update features used by earlier areas.Source: C. Incremental Mapping Without… · p. 4
- The regularizer penalizes changes to active local feature parameters relative to values converged on previous scans, weighted by parameter importance.The weights represent the sensitivity of previous-data loss to parameter changes.Source: C. Incremental Mapping Without… · p. 4
- After each scan converges, importance weights are recalculated, with a cap to prevent gradient explosion and iterations covering the scan's training samples.The recalculation uses the gradient of the loss with respect to each parameter.Source: C. Incremental Mapping Without… · p. 4
- The regularization strength λr adjusts how much historical knowledge is retained, and experiments show the approach reduces the risk of catastrophic forgetting.Source: C. Incremental Mapping Without… · p. 4
IV. EXPERIMENTAL EVALUATION
The evaluation section assesses an incremental, scalable 3D mapping system built from a sparse hierarchical feature grid and neural decoder.
IV. EXPERIMENTAL EVALUATION
- The evaluated system targets incremental and scalable 3D mapping.Source: §IV · p. 4
- Its mapping representation combines a sparse hierarchical feature grid with a neural decoder.Source: §IV · p. 4
- The section analyzes the approach's capabilities and assesses its performance.Source: §IV · p. 4
A. Experimental Setup
The setup evaluates reconstruction on two outdoor LiDAR datasets with near-ground-truth mesh information, then tests incremental scalability on KITTI and presents indoor reconstruction. It also specifies parameter settings and includes MaiCity visual and quantitative comparisons.
A. Experimental Setup
- Two public outdoor LiDAR datasets with near-ground-truth mesh information are used to evaluate mapping accuracy, completeness, and memory efficiency.Source: A. Experimental Setup · p. 4
- MaiCity contains 64 noise-free simulated scans, while Newer College uses handheld LiDAR and has centimeter-level noise and substantial motion distortion.Near-ground-truth data for Newer College comes from a terrestrial scanner.Source: A. Experimental Setup · p. 4
- Table I lists the parameter settings used in the experiments.Source: Table I · p. 4
- The MaiCity figure compares methods using reconstructed meshes and signed-error maps from −5 cm to +5 cm; Table II reports reconstruction-quality metrics.Table II covers completion, accuracy, and Chamfer-L1 in centimeters, plus completion ratio and F-score with a 10 cm threshold.Source: Figure 5 · p. 5, Table II · p. 5
- KITTI odometry is used to validate incremental-mapping scalability, and the experiments also showcase indoor 3D reconstruction.Source: A. Experimental Setup · p. 5
B. Mapping Quality
At a fixed 10 cm voxel size, SHINE-Mapping outperforms the compared methods on the outdoor datasets, including the noisier Newer College scans. The evaluation compares reconstruction quality against TSDF fusion and surface-based systems.
B. Mapping Quality
- The comparison evaluates accuracy and completeness on MaiCity and Newer College against Voxblox, VDB Fusion, and Puma.The authors also evaluate an Ours + DR variant that adds differentiable rendering supervision.Source: B. Mapping Quality · p. 5
- Mapping quality is assessed on meshes reconstructed with marching cubes, using accuracy, completion, Chamfer-L1, completion ratio, and F-score.The reported accuracy uses the ground-truth mesh masked by the intersection area of reconstructions from all compared methods.Source: B. Mapping Quality · p. 5
- At 10 cm, SHINE-Mapping outperforms the compared methods on MaiCity and performs better on noisy Newer College.Table III reports Newer College completion, accuracy, and Chamfer-L1 in cm, plus completion ratio and F-score at a 20 cm threshold.Source: Table III · p. 5, B. Mapping Quality · p. 5
- On Newer College, the method handles measurement noise and removes moving pedestrians while retaining sparse static trees.The paper contrasts this with space carving in Voxblox and VDB Fusion and density thresholding in Puma.Source: B. Mapping Quality · p. 5
C. Memory Efficiency
Across tested voxel sizes, SHINE-Mapping uses less memory while its mapping quality changes little at lower feature-grid resolution. The comparisons relate memory usage to Chamfer-L1 distance.
C. Memory Efficiency
- SHINE-Mapping uses less memory consistently across tested voxel sizes, while its quality hardly worsens at lower feature-grid resolution.Voxblox and VDB Fusion show significantly increased mapping error as resolution decreases.Source: C. Memory Efficiency · pp. 5–6
- The memory-versus-quality comparison covers voxel sizes of 10, 20, 50, and 100 cm, with Chamfer-L1 as the quality measure.For SHINE-Mapping, voxel size denotes the octree leaf-node grid resolution.Source: Figure 7 · p. 6
- Hash-stored features support memory efficiency, whereas Voxblox and VDB Fusion allocate same-sized voxels in truncation bands and ray-carved free space.Source: C. Memory Efficiency · p. 6
D. Scalable Incremental Mapping
SHINE-Mapping incrementally reconstructs a roughly 4 km KITTI driving sequence using regularized updates to hierarchical feature grids. In a qualitative comparison, regularization preserves structures that may vanish or distort without it.
D. Scalable Incremental Mapping
- SHINE-Mapping incrementally reconstructs a KITTI driving sequence spanning about 4 km using regularized updates to hierarchical feature grids.Source: D. Scalable Incremental Mapping · p. 6
- With regularization, structures that may vanish or distort during unregularized incremental mapping are clearly reconstructed.The comparison uses feature leaf nodes with 50 cm resolution.Source: D. Scalable Incremental Mapping · p. 6
E. Indoor Mapping and Filling Occluded Areas
The indoor example uses a finer 3 cm leaf-node resolution and reports detailed reconstruction and reasonable completion in occluded areas. The conclusion describes the broader approach as incremental range-sensor SDF mapping with lower memory than compared methods.
E. Indoor Mapping and Filling Occluded Areas
- Figure 9 shows the reconstructed office floor alongside a close-up mesh and input point cloud, with reasonable completion in the highlighted occluded area.Source: Figure 9 · p. 6
- The indoor reconstruction uses 3 cm leaf nodes and a map memory of 84 MB, detailing furniture and an approximately 1 cm-thick whiteboard while completing occluded areas.The authors use finer resolution than the 10 cm used outdoors to capture indoor details.Source: E. Indoor Mapping and Filling Occluded… · p. 6
V. CONCLUSION
- The conclusion characterizes the method as large-scale 3D SDF mapping from range sensors, with more accurate and complete reconstructions, lower map memory, and incremental online operation.It also reports reasonable guesses about structures in sensor-occluded regions.Source: §V · p. 6