Proxy-GS：面向结构化三维高斯泼溅训练与推理的统一遮挡先验
Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting
用快速代理表示生成遮挡深度图，为高斯泼溅的训练致密化与推理剔除提供统一遮挡先验，可直接迁移到大规模场景的 LOD 与加速渲染管线中。
原文依据 a fast proxy system capable of producing precise occlusion depth maps at a resolution of 1000 x 1000 in under 1 ms
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论三维高斯泼溅（3DGS）已成为实现照片级真实感渲染的高效方法。近期基于 MLP 的变体进一步提升了视觉保真度，但在渲染时引入了大量解码开销。为降低计算成本，已有若干剪枝策略与细节层次（LOD）技术被提出，用于减少大规模场景中的高斯基元数量。然而，我们的分析表明，由于缺乏遮挡感知，仍存在显著的冗余。在本工作中，我们提出 Proxy-GS，一种利用代理表示为高斯引入任意视角遮挡感知的新流程。该方法的核心是一个快速代理系统，能够在 1 毫秒内生成分辨率为 1000 x 1000 的精确遮挡深度图。该代理承担两项作用：其一，引导锚点与高斯的剔除以加速渲染；其二，在训练中引导致密化朝向场景表面，减少遮挡区域的不一致性，从而提升渲染质量。在 MatrixCity Streets 数据集等遮挡严重的场景中，Proxy-GS 相比 Octree-GS 实现超过 2.5 倍的加速，同时提升了渲染质量。
英文原文摘要 1247 字符
3D Gaussian Splatting (3DGS) has emerged as an efficient approach for photorealistic rendering. Recent MLP-based variants further improve visual fidelity but introduce substantial decoding overhead during rendering. To reduce the computational cost, several pruning strategies and level-of-detail (LOD) techniques have been introduced to reduce the number of Gaussian primitives in large-scale scenes. However, our analysis reveals that significant redundancy still remains due to the lack of occlusion awareness. In this work, we propose Proxy-GS, a novel pipeline that uses a proxy representation to introduce occlusion awareness for Gaussians from arbitrary views. At the core of our approach is a fast proxy system capable of producing precise occlusion depth maps at a resolution of 1000 x 1000 in under 1 ms. This proxy serves two roles. First, it guides the culling of anchors and Gaussians to accelerate rendering. Second, it guides densification toward scene surfaces during training, reducing inconsistencies in occluded regions and thereby improving rendering quality. In heavily occluded scenarios such as the MatrixCity Streets dataset, Proxy-GS achieves more than a 2.5x speedup over Octree-GS while also improving rendering quality.