Abstract
Real-time 3D reconstruction from RGB-D sensor data is a cornerstone of robotic mapping and object modeling. While Truncated Signed Distance Functions (TSDF) combined with Marching Cubes are popular, they suffer from significant inaccuracies when dealing with thin structures, often resulting in "exploded" or missing surfaces. We introduce the Directional TSDF (DTSDF), which stores opposing surface data in separate orientation-based channels. Combined with gradient-directed ray casting and a modified marching cubes algorithm, DTSDF achieves superior accuracy and preserves topological consistency for thin features even at coarse voxel resolutions.
The Thinness Problem in Standard TSDF
Standard TSDF methods divide space into a voxel grid where each cell stores the distance to the nearest surface. This approach fails for thin objects—like sheets or thin walls—where the distance values from the front and back surfaces overlap and contradict each other within the same voxel. This interference causes the reconstructed geometry to "blow up," appearing much thicker than reality, or to vanish entirely if the truncation distance is too large relative to the object's thickness.
Methods: Directional Representation and Fusion
The authors propose a three-pronged solution to these geometric artifacts:
- Directional TSDF (DTSDF): Instead of a single scalar, each voxel stores up to six TSDF values corresponding to the $\pm X, \pm Y, \pm Z$ axes. This allows the system to represent multiple surfaces with different orientations within the same spatial volume.
- Gradient-Directed Ray Casting: Rather than simple line-of-sight projection, the system casts rays along the surface normal. This ensures that voxels are updated based on the actual orientation of the surface, drastically reducing errors at corners and steep viewing angles.
- Directional Marching Cubes: The mesh extraction process is overhauled to evaluate all six directional channels. A voting mechanism filters out implausible hypotheses, enabling the extraction of coherent, potentially opposing surfaces from a single mesh unit.
Implementation and Memory Management
To handle the increased data requirements of storing six channels per voxel, the system utilizes a hierarchical hashing structure. The world is divided into blocks, which are indexed via a hash table. Each block contains the directional voxel arrays and corresponding mesh units. This ensures that memory is only allocated for regions containing actual surface data, mitigating the overhead of the multi-channel representation.
Key Results
The DTSDF was evaluated against high-quality models from the Stanford 3D Scanning Repository and real-world indoor datasets.
- Accuracy: In synthetic tests using the "Armadillo" model at a 20mm resolution, DTSDF achieved an RMSE of 3.58 mm, nearly four times more accurate than the state-of-the-art baseline (11.83 mm).
- Detail Preservation: Qualitative analysis shows that thin ridges and sharp features (e.g., on the Dragon model) remain distinct even at low resolutions, whereas standard methods typically merge them into amorphous "blobs."
- Performance: Despite the 2x increase in integration time, the system remains real-time capable, with update frequencies between 13ms and 74ms on mobile hardware.
Conclusion
By explicitly modeling surface orientation, Directional TSDF decouples reconstruction quality from voxel resolution for thin structures. This representation provides a more coherent and accurate mesh for robotic applications where thin features are prevalent. Future research will explore integrating DTSDF into SLAM pipelines to enhance tracking robustness through better frame-to-model registration.
Summary updated