跳转至

文章背景与核心概要

在三维场景编辑领域，如何干净且几何一致地移除多个目标物体一直是一项技术难点。传统的3D高斯泼溅（3DGS）方法在扩展到多物体擦除时，往往难以处理复杂的遮挡关系并维持良好的几何一致性。为此，本文介绍了 CoGeo-GS 框架，它创新性地引入了基于概念驱动的方法，通过语义标签精确定位并隔离目标物体，从而减少前景与背景之间的相互干扰。

为了确保修复区域的高保真度，CoGeo-GS 采用了一套完整的几何感知补全流水线。该流水线有机结合了单目深度先验、基于扩散模型的精细化处理以及边界对齐融合技术，最终实现了卓越的视觉质量和稳定的三维重建效果。该研究已被计算机视觉领域国际会议 ICME 2026 接受。

CoGeo-GS: Concept-Driven and Geometry-Aware Multi-Object Removal in 3D Scenes

Authors: Yuanxiang Ni, Xianliang Huang, Chenhang Ma, Chen Xiao, Yuewen Ma, Ruxin Wang, Hao Zhang
Published: August 27, 2026
Venue: Accepted at ICME 2026
arXiv ID: 2608.26656

Authors: Yuanxiang Ni, Xianliang Huang, Chenhang Ma, Chen Xiao, Yuewen Ma, Ruxin Wang, Hao Zhang
Published: August 27, 2026
Venue: Accepted at ICME 2026
arXiv ID: 2608.26656

摘要

CoGeo-GS 是一个旨在解决3D场景中多物体擦除复杂性的新颖框架。传统的3D高斯泼溅（3DGS）方法在扩展至多物体以及保持几何一致性方面表现不佳，而 CoGeo-GS 引入了概念驱动的方法，利用语义标签来隔离物体。这允许进行精确擦除，并减少了前景与背景元素之间的干扰。为了确保高保真结果，该框架采用了一个几何感知补全流水线，集成了单目深度先验、基于扩散模型的精细化处理和边界对齐融合，从而带来卓越的视觉质量和稳定的重建效果。

Summary

CoGeo-GS is a novel framework designed to address the complexities of multi-object removal in 3D scenes. While traditional 3D Gaussian Splatting (3DGS) methods struggle with scaling to multiple objects and maintaining geometric consistency, CoGeo-GS introduces a concept-driven approach that uses semantic tags to isolate objects. This allows for precise removal and reduced interference between foreground and background elements. To ensure high-fidelity results, the framework employs a geometry-aware completion pipeline that integrates monocular depth priors, diffusion-based refinement, and boundary-aligned blending, resulting in superior visual quality and stable reconstruction.

核心特性

  • 概念感知语义标记：为高斯点分配特定标签，便于在单个优化阶段内进行灵活的多物体选择。
  • 几何感知补全：利用单目深度先验和基于扩散模型的精细化处理，为擦除的区域填补合理且高质量的几何结构。
  • 边界对齐融合：确保擦除区域与原始场景之间的无缝过渡。
  • 几何正则化精细化：在整个编辑过程中稳定整体重建并保持多视角一致性。

Key Features

  • Concept-Aware Semantic Tagging: Assigns specific tags to Gaussians, facilitating flexible, multi-object selection within a single optimization stage.
  • Geometry-Aware Completion: Utilizes monocular depth priors and diffusion-based refinement to fill in removed regions with plausible, high-quality geometry.
  • Boundary-Aligned Blending: Ensures seamless transitions between removed areas and the original scene.
  • Geometry-Regularized Refinement: Stabilizes the overall reconstruction and preserves multi-view consistency throughout the editing process.

技术元数据

属性 详情
主要学科 计算机视觉与模式识别 (cs.CV)
次要学科 人工智能 (cs.AI)
ACM 类别 I.3; I.4
DOI 10.48550/arXiv.2608.26656

Technical Metadata

Attribute Details
Primary Subject Computer Vision and Pattern Recognition (cs.CV)
Secondary Subjects Artificial Intelligence (cs.AI)
ACM Classes I.3; I.4
DOI 10.48550/arXiv.2608.26656

访问论文

Access Paper