Abstract: Mesh plays an indispensable role in dense real-time reconstruction essential in robotics. Efforts have been made to maintain flexible data structures for 3D data fusion, yet an efficient incremental framework specifically designed for online mesh storage and manipulation is missing. We propose a novel framework to compactly generate, update, and refine mesh for scene reconstruction upon a volumetric representation. Maintaining a spatial-hashed field of cubes, we distribute vertices with continuous value on discrete edges that support O(1) vertex accessing and forbid memory redundancy. By introducing Hamming distance in mesh refinement, we further improve the mesh quality regarding the triangle type consistency with a low cost. Lock-based and lock-free operations were applied to avoid thread conflicts in GPU parallel computation. Experiments demonstrate that the mesh memory consumption is significantly reduced while the running speed is kept in the online reconstruction process.
Introduction
In the field of real-time 3D reconstruction—critical for robotics and augmented reality—there is a persistent tension between volumetric data and mesh representations. While volumetric scalar fields like Truncated Signed Distance Fields (TSDF) are excellent for fusing noisy sensor data, they are difficult to render directly. Conversely, meshes are the standard for rendering and physics but are often treated as secondary outputs. This paper introduces a framework that treats the mesh as a first-class citizen within a spatially hashed volumetric grid, optimizing both memory and speed.
The Challenge of Real-Time Meshing
Standard reconstruction methods often produce "loose" meshes with massive amounts of duplicate vertices, necessitating complex auxiliary data structures like KD-trees for neighbor searches. Furthermore, the classic Marching Cubes algorithm frequently generates jagged "cracks" or irregular triangles when surfaces do not align perfectly with the grid. The authors address these inefficiencies by rethinking how mesh data is stored and updated incrementally.
Methodology
The proposed framework utilizes a multi-stage pipeline designed for GPU parallelization, ensuring that the mesh is updated dynamically as new depth and pose data arrive.
Two-Level Spatial Hashing
The system divides 3D space into "Blocks," managed by a hash table for $O(1)$ access. Each block is further subdivided into "Cubes." This structure allows the system to only allocate memory for regions of space that actually contain surfaces, drastically reducing the footprint compared to dense grids.
Compact Edge-Vertex Mapping
A core innovation is the elimination of a separate vertex list. By realizing that every vertex in a Marching Cubes-style mesh must lie on a voxel edge, the researchers store vertices directly on the edges of the cubes. Since adjacent cubes share edges, this naturally prevents duplicate vertices and ensures topological consistency.
Hamming Distance Refinement
To address geometric artifacts, the authors represent a cube's "type" (the configuration of its vertices) as an 8-bit vector. They use Hamming distance to compare the current cube state to a set of "regular" triangle patterns. If noise causes a cube to deviate slightly, the system "snaps" it to the nearest regular pattern. This discrete logic provides a computationally cheap alternative to traditional Laplacian smoothing.
GPU Optimization and Garbage Collection
The pipeline was implemented on a GPU using a lock-based approach with atomic operations, which surprisingly outperformed lock-free alternatives by 2–3x. Additionally, a reference counting garbage collector automatically recycles vertices that are no longer part of any triangle, further optimizing memory usage.
Key Results
Evaluation on standard RGB-D datasets (ICL-NUIM, TUM RGB-D, and large-scale scenes like "burghers") demonstrated significant improvements:
- Memory Efficiency: The method reduced vertex counts by approximately 80% compared to baselines like CHISEL by eliminating redundancy.
- Performance: The system maintains real-time speeds, often processing frames in under 10ms even at high resolutions (8mm cubes).
- Quality: The Hamming distance refinement successfully removed "stair-step" artifacts and cracks, resulting in smoother planes and more consistent triangle shapes.
Conclusion
By integrating mesh storage directly into a spatially hashed volumetric field, this research provides a highly efficient path for real-time reconstruction. While the system currently relies on accurate camera poses and can occasionally filter out fine details, it sets a new standard for memory-efficient online meshing. Future work will focus on integrating this representation into full SLAM systems for massive environment mapping.
Summary updated