DVSH-GS: Dense-Voxel and Spherical Harmonics Guided Gaussian Splatting for Sparse View rendering
IEEE Signal Processing Letters · 2026 · 33 · 2805-2809
DOI: 10.1109/lsp.2026.3707490 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Fansen Zhou und 2 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Gefördert von: Local Scientific and Technological Development Funding Projects for Central Guidance in Sichuan, Chengdu Science and Technology Program, Police Integration Computing Key Laboratory of Sichuan Province.
Abstract (Originalsprache)
3D Gaussian Splatting (3DGS) has recently emerged as an efficient explicit representation for high-quality novel view synthesis. However, under sparse-view supervision, 3DGS suffers from severe overfitting, resulting in distorted geometry and poor generalization to unseen viewpoints. Current dropout-based regularization techniques attempt to mitigate this by randomly discarding individual Gaussians, yet their efficacy is fundamentally hindered by the strong spatial redundancy among neighboring Gaussians, which allows for trivial compensation effects. In this letter, we identify two primary drivers of sparse-view overfitting: local Gaussian redundancy and over-parameterization in high-degree spherical harmonics (SH). To address these, we propose DVSH-GS, a dual-regularization framework. First, we introduce a structured density-guided spatial dropout mechanism that removes spatially correlated Gaussians groups (voxels) rather than isolated points, thereby explicitly breaking local dependencies. Second, we present a selective SH dropout strategy that suppresses high-degree SH coefficients during training to prevent view-dependent appearance overfitting. Extensive experiments on multiple benchmarks demonstrate that DVSH-GS outperforms existing methods in sparse-view scenarios while remaining a plug-and-play solution with minor implementation complexity.
Förderung
- Local Scientific and Technological Development Funding Projects for Central Guidance in Sichuan
- Chengdu Science and Technology Program
- Police Integration Computing Key Laboratory of Sichuan Province
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Zhou, F., Zhang, F., & Chen, H. (2026). DVSH-GS: Dense-Voxel and Spherical Harmonics Guided Gaussian Splatting for Sparse View rendering. IEEE Signal Processing Letters, 33, 2805-2809. https://doi.org/10.1109/lsp.2026.3707490
Vancouver
Zhou F, Zhang F, Chen H. DVSH-GS: Dense-Voxel and Spherical Harmonics Guided Gaussian Splatting for Sparse View rendering. IEEE Signal Processing Letters. 2026;33:2805-2809. doi:10.1109/lsp.2026.3707490
BibTeX
@article{Zhou2026,
author = {Zhou, Fansen and Zhang, Fan and Chen, Honggang},
title = {DVSH-GS: Dense-Voxel and Spherical Harmonics Guided Gaussian Splatting for Sparse View rendering},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2805--2809},
doi = {10.1109/lsp.2026.3707490}
}
EUVIMED