Research map / 3D vision / Gaussian splatting
Fully Explicit Dynamic 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 Ex4DGS, a fully explicit dynamic 3D Gaussian Splatting method that separates static and dynamic Gaussians, interpolates keyframed positions (cubic Hermite), rotations (Slerp) and opacity (two-Gaussian mixture), and uses progressive training and point backtracking pruning.
Problem
Dynamic novel view synthesis with Gaussian splatting is slow and memory-heavy, and existing 4D Gaussian methods rely on dense point clouds and implicit networks, limiting fast rendering and sparse-initialization robustness.
Approach
Store explicit keyframe positions and rotations only for dynamic Gaussians (static ones move linearly), interpolate them with cubic Hermite and Slerp, model temporal opacity with a two-Gaussian mixture, and grow the time window progressively while pruning by accumulated backtracked error.
Key results
- On Neural 3D Video with sparse COLMAP input, Ex4DGS reaches 32.11 average PSNR versus 31.67 for 3DGStream and 30.64 for STG, at 120.6 FPS and 115 MB.
- On Technicolor, Ex4DGS achieves 33.62 PSNR and 0.088 LPIPS, outperforming STG (33.23 PSNR) on PSNR.
- Ex4DGS renders at 62 FPS on a single 2080Ti GPU for 300-frame 1352x1014 scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ex4DGS instead of 3DGStream with sparse COLMAP initialization | Dynamic novel view synthesis · Neural 3D Video | Average PSNR (dB): 31.67 -> 32.11 | Helped |
| Replace cubic Hermite position interpolation with linear interpolation (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.12 | Hurt |
| Remove temporal opacity modeling (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.42 | Hurt |
| Replace Slerp rotation interpolation with linear interpolation (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.26 | Hurt |
| Remove progressive growing (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.02 | Hurt |
| Replace position and rotation interpolation with linear interpolation (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.32 | Hurt |
| Remove motion regularization (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.37 | Hurt |
| Remove dynamic point extraction (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 28.58 | Hurt |
| Remove point backtracking (ablation) | Ex4DGS · Ablation study | PSNR: 32.11 -> 31.40 | Hurt |
| Set dynamic point conversion rate to 2% (ablation) | Ex4DGS · Cook Spinach | PSNR: 32.36 -> 33.04 | Helped |
| Use keyframe interval 10 rather than smaller intervals (ablation) | Ex4DGS · Cook Spinach, skipped frames 1 | PSNR: 31.17 -> 33.04 | Helped |
| Add dynamic Gaussians to handle stationary objects whose colors change | 3DGS · Coffee Martini | PSNR: 21.69 -> 26.07 | Helped |
Surprising findings
- Concurrent explicit dynamic Gaussian methods (4DGS, 4DGaussians, STG) yield unsatisfactory results when initialized with sparse first-frame point clouds.
- Color-only changes (e.g., shadows, reflections) cause large quality loss if not modeled as dynamic points.
Limitations
- Gets stuck in local minima for newly appeared objects not initialized with 3D points.
- Training on monocular videos is challenging due to scale ambiguity; all Gaussians become dynamic.
Experimental setup
Datasets: Neural 3D Video dataset Technicolor dataset
Benchmarks: Neural 3D Video dataset Technicolor dataset
Metrics: PSNR SSIM LPIPS FPS training time model size
Baselines: NeRFPlayer HyperReel DyNeRF HexPlane K-Planes MixVoxels Im4D 4K4D STG 4DGS 4DGaussians 3DGStream
Builds on: 3D Gaussian Splatting Mip-Splatting Spherical Linear Interpolation cubic Hermite spline
Closest papers by meaning
- 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering (NeurIPS 2025 · similarity 0.94)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.93)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.93)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.93)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
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.