Scholarlink

<p>In recent years, 3D Gaussian Splatting has become a mainstream method in 3D reconstruction, with its Gaussian ellipsoid representation widely adopted. However, it suffers from excessive memory consumption and numerous low-contribution redundant Gaussian primitives that degrade rendering quality. Reducing primitive storage while preserving rendering quality remains a fundamental challenge, as existing methods relying on direct pruning often cause unstable quality degradation. To address this, we propose a dynamic hybrid optimization strategy (DHO-3DGS) to enhance the efficiency of Gaussian primitive representation. Specifically: (1) A dynamic weight optimization strategy accelerates geometric initialization in the densification stage; (2) a dynamic polarization loss term is introduced in the pruning stage to impose polarization constraints on Gaussian opacity, combined with transparency pruning for precise removal of redundant primitives; (3) an edge detection operator is integrated in the fine-tuning stage to strengthen edge structure expression in 3D geometries and avoid detail blurring. Experimental results on Deep Blending, Tanks &amp; Temples, and MipNeRF360 datasets show that DHO-3DGS outperforms state-of-the-art (SOTA) methods, achieving an average 0.12dB improvement in PSNR, 20% reduction in memory usage, and 48% compression for the high-compression variant while maintaining equivalent rendering quality. Additionally, our method exhibits strong portability, validated on original 3DGS, Taming 3DGS, and PUP 3DGS. Code is available at: <a href="https://github.com/lllzt47/DHO-3DGS.">https://github.com/lllzt47/DHO-3DGS.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DHO-3DGS: 3D Gaussian Splatting with dynamic hybrid optimization

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Haixiang Zhang,
  • Zhentao Lv,
  • Shangfei Zheng

摘要

In recent years, 3D Gaussian Splatting has become a mainstream method in 3D reconstruction, with its Gaussian ellipsoid representation widely adopted. However, it suffers from excessive memory consumption and numerous low-contribution redundant Gaussian primitives that degrade rendering quality. Reducing primitive storage while preserving rendering quality remains a fundamental challenge, as existing methods relying on direct pruning often cause unstable quality degradation. To address this, we propose a dynamic hybrid optimization strategy (DHO-3DGS) to enhance the efficiency of Gaussian primitive representation. Specifically: (1) A dynamic weight optimization strategy accelerates geometric initialization in the densification stage; (2) a dynamic polarization loss term is introduced in the pruning stage to impose polarization constraints on Gaussian opacity, combined with transparency pruning for precise removal of redundant primitives; (3) an edge detection operator is integrated in the fine-tuning stage to strengthen edge structure expression in 3D geometries and avoid detail blurring. Experimental results on Deep Blending, Tanks & Temples, and MipNeRF360 datasets show that DHO-3DGS outperforms state-of-the-art (SOTA) methods, achieving an average 0.12dB improvement in PSNR, 20% reduction in memory usage, and 48% compression for the high-compression variant while maintaining equivalent rendering quality. Additionally, our method exhibits strong portability, validated on original 3DGS, Taming 3DGS, and PUP 3DGS. Code is available at: https://github.com/lllzt47/DHO-3DGS.