Research map / 3D vision / Gaussian splatting
MoDec-GS: Global-to-Local Motion Decomposition and Temporal Interval Adjustment for Compact Dynamic 3D Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoDec-GS, a compact monocular dynamic 3D Gaussian splatting framework that decomposes motion into global anchor deformation and local Gaussian deformation with an automatically adjusted temporal interval per local canonical scaffold.
Problem
Dynamic 3D Gaussian splatting from monocular video has high storage demands and struggles to represent long-duration content with complex combinations of global and local motion.
Approach
Global-to-Local Motion Decomposition: anchor-based (Scaffold-GS) global canonical scaffolds are deformed by a tiny global hexplane into time-segment local scaffolds, which then get explicit local Gaussian deformation via a shared hexplane; Temporal Interval Adjustment shrinks segments with high accumulated positional gradients, without optical flow.
Key results
- On the iPhone dataset, MoDec-GS averages 14.60 mPSNR with 18.37 MB storage versus SC-GS's 13.90 mPSNR at 232.4 MB.
- On HyperNeRF, MoDec-GS reaches 27.78 PSNR at 40.82 MB, beating 4DGS (27.44 PSNR, 72.65 MB).
- On Nvidia monocular, MoDec-GS reaches 26.65 PSNR at 39.64 MB versus 4DGS's 25.82 PSNR at 67.44 MB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoDec-GS instead of SC-GS | dynamic 3D Gaussian reconstruction · iPhone (Average) | mPSNR: 13.90 -> 14.60 | Helped |
| Use MoDec-GS instead of SC-GS | dynamic 3D Gaussian reconstruction · iPhone (Average) | Storage: 232.4 -> 18.37 | Helped |
| Use MoDec-GS instead of 4DGS | dynamic 3D Gaussian reconstruction · HyperNeRF | PSNR: 27.44 -> 27.78 | Helped |
| Use MoDec-GS instead of SC-GS | dynamic 3D Gaussian reconstruction · HyperNeRF | LPIPS: 0.213 -> 0.219 | Hurt |
| Use MoDec-GS instead of 4DGS | dynamic 3D Gaussian reconstruction · Nvidia monocular | PSNR: 25.82 -> 26.65 | Helped |
| Replace single-stage Gaussian deformation with single-stage anchor deformation (ablation) | MoDec-GS · iPhone (Average) | mPSNR: 13.73 -> 13.56 | Hurt |
| Replace single-stage Gaussian deformation with single-stage anchor deformation (ablation) | MoDec-GS · iPhone (Average) | Storage: 78.54 -> 36.92 | Helped |
| Add a global hexplane to configure Local CS (ablation) | single-stage anchor deformation · iPhone (Average) | mPSNR: 13.73 -> 13.93 | Helped |
| Add LGD after GAD (ablation) | two-stage MoDec-GS · iPhone (Average) | mPSNR: 13.93 -> 14.48 | Helped |
| Use LGD after GAD instead of anchor deformation in both stages (ablation) | two-stage MoDec-GS · iPhone (Average) | Storage: 55.29 -> 49.70 | Helped |
| Use a smaller hexplane (ablation) | GLMD · iPhone (Average) | mPSNR: 14.48 -> 14.46 | Hurt |
| Use a smaller hexplane (ablation) | GLMD · iPhone (Average) | Storage: 49.70 -> 22.67 | Helped |
Surprising findings
- Anchor-based deformation alone slightly reduces rendering quality relative to explicit Gaussian deformation but cuts storage by about 52%.
Limitations
- Relies on COLMAP-generated initial point clouds manually produced per dataset.
- Temporal interval adjustment and many hyperparameters are set empirically.
Experimental setup
Datasets: iPhone (DyCheck) HyperNeRF Nvidia monocular
Benchmarks: iPhone (DyCheck) HyperNeRF Nvidia monocular
Metrics: PSNR SSIM LPIPS storage (MB)
Baselines: SC-GS Deformable 3DGS 4DGS
Builds on: 4DGS Scaffold-GS Deformable 3DGS
Closest papers by meaning
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.93)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.92)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.92)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- Orientation-anchored Hyper-Gaussian for 4D Reconstruction from Casual Videos (NeurIPS 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.