HiCo-GS: New Framework Enhances Octree Gaussian Splatting for City-Scale Rendering
A recent paper available on arXiv (2608.14136) presents HiCo-GS, a framework for high-fidelity reconstruction aimed at overcoming the challenges associated with octree-based anchor Gaussian Splatting for city-scale novel view synthesis. This approach focuses on the issue of cross-level feature isolation, where independently optimized multi-level anchor features can result in color discrepancies on building facades and excessive smoothing in textured areas. HiCo-GS integrates two key modules: Cross-Level Context Aggregation (CLCA) and Geometric Consistency. CLCA utilizes the octree's spatial structure to combine context vectors into parent-self-child triplets, processed through a lightweight MLP with a residual connection, enhancing hierarchical prior injection. The paper is noted as a cross submission, suggesting potential conference presentation, and it holds significance for digital art and technology, especially in crafting realistic virtual environments.
Key facts
- Paper arXiv:2608.14136 introduces HiCo-GS framework.
- HiCo-GS addresses cross-level feature isolation in octree-based anchor Gaussian Splatting.
- Problem causes color drift on building facades and over-smoothing in textured regions.
- Two modules: Cross-Level Context Aggregation (CLCA) and Geometric Consistency (implied).
- CLCA uses octree spatial containment to create parent-self-child triplets.
- Triplets fused via lightweight MLP with residual connection.
- Coarse-level structural priors flow down to fine-level anchors.
- Announcement type: cross, suggesting conference submission.
Entities
Institutions
- arXiv