MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting

Sangwoon Kwak, Joonsoo Kim, Jun Young Jeong, Won-Sik Cheong, Jihyong Oh, Munchurl Kim

Introduction

3D Gaussian Splatting (3DGS) has set new standards for high-fidelity, real-time neural rendering of static scenes. However, extending this efficiency to dynamic environments—characterized by complex, multi-scale motions—remains a significant challenge. Current dynamic 3DGS methods often suffer from prohibitive storage requirements due to the need to store 4D attributes for millions of Gaussians. Moreover, they frequently struggle to accurately capture both coarse global movements and fine-grained local deformations, leading to visual artifacts like ghosting or blurring.

MoDec-GS addresses these limitations through a novel "divide and conquer" strategy. By introducing Global-to-Local Motion Decomposition (GLMD) and Temporal Interval Adjustment (TIA), the framework achieves state-of-the-art rendering quality while drastically reducing model size.

Methodology: GLMD and TIA

The core of MoDec-GS is the Global-to-Local Motion Decomposition (GLMD) pipeline, which separates motion into two distinct stages to optimize representational efficiency.

Overview of the MoDec-GS architecture, highlighting the Global-to-Local Motion Decomposition (GLMD) and Temporal Interval Adjustment (TIA) modules.

Global-to-Local Motion Decomposition (GLMD)

GLMD builds upon the anchor-based Scaffold-GS representation. It decomposes motion into:

  1. Global Anchor Deformation (GAD): A compact global "hexplane" deforms the underlying anchors. This stage captures broad, scene-wide movements efficiently, as moving a single anchor is computationally cheaper than moving all associated Gaussians.
  2. Local Gaussian Deformation (LGD): A shared local hexplane refines the individual Gaussians spawned by the anchors, capturing high-frequency details and subtle deformations.

A learnable masking strategy ($d_G, d_L$) is employed to identify static background elements, allowing the model to bypass deformation logic where unnecessary, further conserving computation and storage.

Temporal Interval Adjustment (TIA)

Standard dynamic models typically utilize fixed temporal segments. MoDec-GS introduces Temporal Interval Adjustment (TIA) to handle non-uniform motion intensity. During training, TIA monitors positional gradients; segments with high gradients (indicating rapid or complex motion) are automatically assigned narrower temporal windows. This adaptive approach provides higher temporal resolution exactly where it is needed most.

Experiments and Results

The performance of MoDec-GS was evaluated on three major monocular video benchmarks: Dycheck-iPhone, HyperNeRF, and the Nvidia Monocular Dataset.

Qualitative and quantitative comparison of MoDec-GS against state-of-the-art methods. MoDec-GS preserves sharp details in complex motions while maintaining a significantly smaller storage footprint.

Quantitative Performance

MoDec-GS consistently outperformed existing methods like SC-GS and 4DGS. On the Dycheck-iPhone dataset, it achieved a +0.7dB PSNR improvement while reducing storage by 94% compared to SC-GS. Across all tested datasets, the model maintained an average storage reduction of 70% relative to its closest competitors.

Qualitative Analysis

As shown in Figure 1, MoDec-GS excels at rendering complex movements, such as a knife slicing through a lemon. While competing methods produce blurry results or "ghosting" artifacts, MoDec-GS preserves sharp edges and fine textures. This superior quality is attributed to the GLMD's ability to resolve multi-scale motion and TIA's adaptive temporal resolution.

Conclusion

MoDec-GS represents a significant advancement in dynamic 3D Gaussian Splatting. By decomposing motion into global and local components and adaptively adjusting temporal intervals, it provides a highly compact yet high-fidelity solution for dynamic scene reconstruction. While current limitations include a dependency on COLMAP for initialization, future work focusing on "on-the-fly" initialization and advanced feature quantization could further enhance the framework's efficiency and accessibility.

Summary updated

Abstract

3D Gaussian Splatting (3DGS) has made significant strides in scene representation and neural rendering, with intense efforts focused on adapting it for dynamic scenes. Despite delivering remarkable rendering quality and speed, existing methods struggle with storage demands and representing complex real-world motions. To tackle these issues, we propose MoDecGS, a memory-efficient Gaussian splatting framework designed for reconstructing novel views in challenging scenarios with complex motions. We introduce GlobaltoLocal Motion Decomposition (GLMD) to effectively capture dynamic motions in a coarsetofine manner. This approach leverages Global Canonical Scaffolds (Global CS) and Local Canonical Scaffolds (Local CS), extending static Scaffold representation to dynamic video reconstruction. For Global CS, we propose Global Anchor Deformation (GAD) to efficiently represent global dynamics along complex motions, by directly deforming the implicit Scaffold attributes which are anchor position, offset, and local context features. Next, we finely adjust local motions via the Local Gaussian Deformation (LGD) of Local CS explicitly. Additionally, we introduce Temporal Interval Adjustment (TIA) to automatically control the temporal coverage of each Local CS during training, allowing MoDecGS to find optimal interval assignments based on the specified number of temporal segments. Extensive evaluations demonstrate that MoDecGS achieves an average 70% reduction in model size over stateoftheart methods for dynamic 3D Gaussians from realworld dynamic videos while maintaining or even improving rendering quality.
BibTeX
@inproceedings{kwak2025modec,
  author = {Sangwoon Kwak and Joonsoo Kim and Jun Young Jeong and Won-Sik Cheong and Jihyong Oh and Munchurl Kim},
  title = {{MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.01059},
}
Figure