Shell-Guided Compression of Voxel Radiance Fields.

Area of Science:

• Computer Vision
• 3D Graphics
• Machine Learning

Background:

• Large-scale voxel-based models in 3D reconstruction suffer from high memory consumption and redundant data.
• Existing methods often struggle to balance storage efficiency with rendering accuracy.

Purpose of the Study:

• To introduce a novel compression method, Shell-guided compression of Voxel Radiance Fields (SVRF), for optimizing voxel-based models.
• To significantly reduce storage costs while preserving rendering quality in 3D reconstruction.

Main Methods:

• Implemented a Shell-like Constraint to prioritize surface-neighboring voxels and eliminate redundant ones.
• Utilized Adaptive Thresholds and a Dynamic Pruning Strategy for precise and scene-appropriate voxel removal.
• Employed self-supervised learning guided by predicted depth, without requiring additional labels.

Main Results:

• Achieved over 70% compression of voxel grids.
• Maintained comparable rendering quality to original models.
• Demonstrated seamless integration with existing voxel-grid-based methods.

Conclusions:

• SVRF offers an effective solution for memory and storage challenges in large-scale voxel models.
• The method is efficient, accurate, and broadly applicable to various 3D reconstruction tasks.
• The proposed approach facilitates practical deployment of complex 3D models.