Scholarlink

<p>3D Gaussian splatting (3DGS) has emerged as a powerful explicit representation for real-world scene reconstruction. Nevertheless, existing feed-forward generalizable 3DGS methods often suffer from depth estimation errors in scenes with weak textures and complex geometries, which further introduce noticeable artifacts and distortions in Gaussian rendering. In this work, we present a generalizable 3DGS framework that integrates reference-preserving cross-view feature aggregation with geometry-frequency coordinated refinement. Specifically, the cross-view attention (CVA) module keeps the reference-view representation stable while conditioning source-view features on reference-view context, leading to more reliable depth probability estimation before Gaussian mean initialization. To further improve reconstruction quality for challenging scenes, the bidirectional coordination optimization (BCO) pipeline connects feed-forward Gaussian initialization with compact anchor-based neural Gaussian optimization. BCO regularizes Gaussian distributions through geometric support constraints and progressively emphasizes high-frequency regions through edge-aware frequency supervision. Experiments on DTU, LLFF, Tanks and Temples, and our custom real-world scenes show consistent improvements over representative baselines in PSNR, SSIM, LPIPS, and memory usage under the reported evaluation settings. Our code and dataset are publicly available at <a href="https://github.com/lililihu8/GeneralizableGS.git">https://github.com/lililihu8/GeneralizableGS.git</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

Generalizable 3D Gaussian splatting with cross-view attention and frequency-aware bidirectional coordination

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Chao Li,
  • Wenjie Na,
  • Chenpeng Yao,
  • Chengju Liu,
  • Qijun Chen

摘要

3D Gaussian splatting (3DGS) has emerged as a powerful explicit representation for real-world scene reconstruction. Nevertheless, existing feed-forward generalizable 3DGS methods often suffer from depth estimation errors in scenes with weak textures and complex geometries, which further introduce noticeable artifacts and distortions in Gaussian rendering. In this work, we present a generalizable 3DGS framework that integrates reference-preserving cross-view feature aggregation with geometry-frequency coordinated refinement. Specifically, the cross-view attention (CVA) module keeps the reference-view representation stable while conditioning source-view features on reference-view context, leading to more reliable depth probability estimation before Gaussian mean initialization. To further improve reconstruction quality for challenging scenes, the bidirectional coordination optimization (BCO) pipeline connects feed-forward Gaussian initialization with compact anchor-based neural Gaussian optimization. BCO regularizes Gaussian distributions through geometric support constraints and progressively emphasizes high-frequency regions through edge-aware frequency supervision. Experiments on DTU, LLFF, Tanks and Temples, and our custom real-world scenes show consistent improvements over representative baselines in PSNR, SSIM, LPIPS, and memory usage under the reported evaluation settings. Our code and dataset are publicly available at https://github.com/lililihu8/GeneralizableGS.git.