Title: Moving Past Static Gaussians: An Empirical Study of Architectural Frameworks for Dynamic 3D Scene Reconstruction

Abstract:

While 3D Gaussian Splatting (3DGS) has become a prominent technique for capturing evolving environments, it remains essential to understand the trade-offs inherent in various methodological approaches. This study offers a thorough examination of dynamic 3DGS strategies, dividing them into two primary categories: structure-guided techniques, which utilize auxiliary representations such as canonical spaces, grids, or deformation fields to account for temporal variations, and gaussian-centric approaches that embed dynamic properties directly into the primitives through continuous functions or 4D representations. By assessing leading methods from both categories on the D-NeRF benchmark, we identify a distinct dichotomy in performance. Structure-guided models excel in producing higher-fidelity reconstructions and maintaining more compact model sizes. In contrast, gaussian-centric methods offer substantially faster rendering capabilities suitable for real-time applications, albeit with increased variability in visual quality and potentially higher storage demands. These results underscore a core compromise between reconstruction accuracy and model compactness on one hand, and rendering velocity on the other, offering valuable guidance for advancing future research and practical implementations in dynamic scene reconstruction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC