高效模型与推理加速 · CVPR2026 ·2026.06.03

SpeeDe3DGS:基于时序剪枝与运动分组的快速可形变三维高斯泼溅

SpeeDe3DGS: Speedy Deformable 3D Gaussian Splatting with Temporal Pruning and Motion Grouping

Allen Tu、Haiyang Ying、Alex Hanson、Yonghan Lee、Tom Goldstein、Matthias Zwicker

与你有关的部分

对做动态 3DGS 的人,这篇给出了一套把逐高斯神经推理换成共享 SE(3) 群变换 + 时序敏感剪枝的加速范式,可直接迁移到任何带神经运动场的可形变场景表示上,在几乎不掉画质的前提下把渲染提速一个数量级。

原文依据 accelerates rendering by 6.78x on average while maintaining neural-field fidelity and using 10x fewer primitives
可迁移到 动态场景新视角合成可形变 3D 重建实时渲染

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)的动态扩展通过神经运动场实现了高质量重建,但逐高斯的神经推理使这类模型计算开销高昂。我们在 DeformableGS 的基础上提出快速可形变三维高斯泼溅(SpeeDe3DGS),通过三个互补模块弥合效率与保真度之间的差距:时序敏感性剪枝(TSP)借助时序聚合的敏感性分析移除低影响高斯;时序敏感性采样(TSS)对时间戳施加扰动以抑制漂浮伪影并提升时序一致性;GroupFlow 将学习到的形变场蒸馏为共享的 SE(3) 变换,以实现高效的组级运动。在 MonoDyGauBench 的 50 个动态场景上,将 TSP 与 TSS 集成进 DeformableGS 后,渲染平均加速 6.78 倍,同时保持神经场保真度,且所用基元数量减少 10 倍。进一步加入 GroupFlow 后,渲染速度提升至 13.71 倍,训练时间缩短 2.53 倍,在保持更优图像质量的同时速度超越所有基线。

英文原文摘要 1068 字符

Dynamic extensions of 3D Gaussian Splatting (3DGS) achieve high-quality reconstructions through neural motion fields, but per-Gaussian neural inference makes these models computationally expensive. Building on DeformableGS, we introduce Speedy Deformable 3D Gaussian Splatting (SpeeDe3DGS), which bridges this efficiency-fidelity gap through three complementary modules: Temporal Sensitivity Pruning (TSP) removes low-impact Gaussians via temporally aggregated sensitivity analysis, Temporal Sensitivity Sampling (TSS) perturbs timestamps to suppress floaters and improve temporal coherence, and GroupFlow distills the learned deformation field into shared SE(3) transformations for efficient groupwise motion. On the 50 dynamic scenes in MonoDyGauBench, integrating TSP and TSS into DeformableGS accelerates rendering by 6.78x on average while maintaining neural-field fidelity and using 10x fewer primitives. Adding GroupFlow culminates in 13.71x faster rendering and 2.53x shorter training, surpassing all baselines in speed while preserving superior image quality.

标签

会议 CVPR2026
页码 26083-26093
Bibkey
来源 CVF Open Access