LG-HCC: Local Geometry-Aware Hierarchical Context Compression for 3D Gaussian Splatting
Summary: arXiv:2603.28431v3 Announce Type: replace-cross
Abstract
While 3D Gaussian Splatting (3DGS) offers high-fidelity real-time rendering capabilities, the considerable storage requirements it demands pose significant challenges for practical implementation. Recent advancements in anchor-based 3DGS compression techniques have made strides in reducing Gaussian redundancy through sophisticated context models. However, these methods often neglect to account for explicit geometric dependencies, which can lead to structural degradation and suboptimal rate-distortion performance.
Introduction
In this paper, we introduce a novel framework termed Local Geometry-aware Hierarchical Context Compression (LG-HCC) designed specifically for 3DGS. The core innovation of LG-HCC lies in its integration of inter-anchor geometric correlations into the processes of anchor pruning and entropy coding, facilitating a more compact representation of 3D data.
Main Contributions
- Neighborhood-Aware Anchor Pruning (NAAP): This strategy assesses the importance of anchors through weighted neighborhood feature aggregation. By merging low-contribution anchors with salient neighbors, we create a more compact and geometry-consistent anchor set.
- Hierarchical Entropy Coding: We introduce a hierarchical coding scheme that leverages coarse-to-fine priors. This is achieved through a lightweight Geometry-Guided Convolution (GG-Conv) operator, allowing for spatially adaptive context modeling and rate-distortion optimization.
Experimental Results
Comprehensive experiments have been conducted to evaluate the performance of LG-HCC. The results demonstrate that our approach effectively addresses structural preservation issues commonly encountered in traditional models. Notably, LG-HCC achieves superior geometric integrity and rendering fidelity, with storage savings reaching up to 30.85 times compared to the Scaffold-GS baseline on the widely recognized Mip-NeRF360 dataset.
Conclusion
In summary, the LG-HCC framework represents a significant advancement in the field of 3D Gaussian Splatting compression. By incorporating local geometric awareness into the compression process, LG-HCC not only enhances the quality of rendered images but also reduces the storage burden associated with 3D data representation. The findings of this study pave the way for more efficient and practical applications of 3DGS in real-time rendering scenarios.
Future Work
Looking ahead, further exploration into optimizing the computational efficiency of the GG-Conv operator and expanding the applicability of LG-HCC to various rendering tasks will be a focus. Additionally, testing the framework on diverse datasets could provide more insights into its versatility and robustness in handling complex geometric structures.