Morton Code-Based Geometry-Adaptive Surface Reconstruction.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Geometric Deep Learning

Background:

• Neural implicit surface representations achieve notable 3D reconstruction results.
• Existing methods struggle with noise in smooth areas and detail loss in complex regions due to inadequate spatial structure modeling.

Purpose of the Study:

• To develop a geometry-adaptive surface reconstruction method that addresses limitations of current neural implicit techniques.
• To leverage Morton codes for explicit spatial structure modeling in 3D reconstruction.

Main Methods:

• Proposed a novel method utilizing Morton codes to map 3D space onto octree traversal paths, creating a spatial structural prior.
• Implemented an implicit octree that generates unique root-to-leaf trajectories for query points.
• Developed spatially adaptive weights to modulate multi-resolution geometric features, prioritizing low-frequency features in flat regions and high-frequency features in detailed areas.

Main Results:

• The proposed method demonstrates competitive performance across various datasets.
• Achieved superior reconstruction of sharp features and intricate geometric details compared to existing approaches.
• Effectively suppressed noise in smooth regions while preserving fine details in complex areas.

Conclusions:

• The geometry-adaptive surface reconstruction method based on Morton codes offers a robust solution for high-fidelity 3D reconstruction.
• Explicit spatial structure modeling via octree traversal and adaptive feature modulation is crucial for overcoming limitations in neural implicit surfaces.
• This approach advances the state-of-the-art in reconstructing complex and detailed 3D geometries.