Scholarlink

<p>3D Gaussian Splatting (3DGS) has recently emerged as an efficient and high-quality paradigm for 3D reconstruction. Nevertheless, it frequently produces floating artifacts—isolated structures that significantly impair visual quality. While these artifacts are notably persistent in cases of low-quality initialization, their fundamental cause remains insufficiently explored. In this work, we revisit this problem from a frequency-inspired perspective and argue that under-optimized Gaussians are a primary factor behind floating artifact formation. Based on this insight, we introduce Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS), a simple yet effective framework that selectively enlarges under-optimized Gaussians to enhance low-frequency learning, while employing depth-aware and scale-aware adaptive control to retain high-frequency details. Extensive experiments across multiple datasets demonstrate that EFA-GS effectively suppresses floating artifacts without compromising fine structures. Specifically, our method achieves a 1.68 dB PSNR improvement over the baseline on our RWLQ dataset and demonstrates robust performance in downstream 3D editing tasks. Our implementation is available at <a href="https://jcwang-gh.github.io/EFA-GS.">https://jcwang-gh.github.io/EFA-GS.</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

Frequency-aware optimization for floating artifact removal in 3D Gaussian splatting

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Jianchao Wang,
  • Peng Zhou,
  • Cen Li,
  • Rong Quan,
  • Jie Qin

摘要

3D Gaussian Splatting (3DGS) has recently emerged as an efficient and high-quality paradigm for 3D reconstruction. Nevertheless, it frequently produces floating artifacts—isolated structures that significantly impair visual quality. While these artifacts are notably persistent in cases of low-quality initialization, their fundamental cause remains insufficiently explored. In this work, we revisit this problem from a frequency-inspired perspective and argue that under-optimized Gaussians are a primary factor behind floating artifact formation. Based on this insight, we introduce Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS), a simple yet effective framework that selectively enlarges under-optimized Gaussians to enhance low-frequency learning, while employing depth-aware and scale-aware adaptive control to retain high-frequency details. Extensive experiments across multiple datasets demonstrate that EFA-GS effectively suppresses floating artifacts without compromising fine structures. Specifically, our method achieves a 1.68 dB PSNR improvement over the baseline on our RWLQ dataset and demonstrates robust performance in downstream 3D editing tasks. Our implementation is available at https://jcwang-gh.github.io/EFA-GS.