Research map / 3D vision / Gaussian splatting
Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene
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.
Swift4D splits 3D Gaussians into static and dynamic sets via a learnable dynamic parameter and applies a compact 4D hash-grid deformation decoder only to dynamic primitives, achieving fast, compact multi-view dynamic scene reconstruction.
Problem
Dynamic novel view synthesis with NeRF or 3D Gaussian Splatting methods is slow to train and storage-heavy, and uniformly modeling static and dynamic Gaussians wastes computation and blurs dynamic regions.
Approach
Learn a per-Gaussian dynamic parameter supervised by 2D temporal-variance masks, then deform only the dynamic Gaussians using a multi-resolution 4D hash encoder and multi-head MLP decoder, combined with temporal importance pruning instead of opacity reset.
Key results
- On N3DV, Swift4D reaches 32.23 PSNR and 0.043 LPIPS with 25 min training, 120 MB, and 125 FPS; Swift4D Lite reaches 31.79 PSNR in 20 min with 30 MB.
- On MeetRoom, Swift4D reports 32.05 PSNR, 0.3 hours training, and 40 MB storage, versus 3DGStream's 30.79 PSNR, 0.6 hours, and 1230 MB.
- The method claims at least 20x faster training than methods with similar rendering quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Swift4D instead of SpaceTimeGS | dynamic scene reconstruction on N3DV · N3DV | PSNR: 32.05 -> 32.23 | Helped |
| use Swift4D instead of SpaceTimeGS | dynamic scene reconstruction on N3DV · N3DV | LPIPS: 0.044 -> 0.043 | Helped |
| use Swift4D instead of 3DGStream | dynamic scene reconstruction · MeetRoom | PSNR: 30.79 -> 32.05 | Helped |
| use Swift4D instead of 3DGStream | dynamic scene reconstruction · MeetRoom | size: 1230 -> 40 | Helped |
| add dynamic-static decomposition (ablation) | Swift4D · sear steak and flame steak | PSNR: 32.68 -> 33.83 | Helped |
| add dynamic-static decomposition (ablation) | Swift4D · sear steak and flame steak | training time: 35 -> 25 | Helped |
| replace the 4D Hash encoder with multi-planes (ablation) | Swift4D · sear steak and flame steak | PSNR: 33.83 -> 33.48 | Hurt |
| replace the 4D Hash encoder with multi-planes (ablation) | Swift4D · sear steak and flame steak | training time: 25 -> 28 | Hurt |
| use hash table size 2^15 and lambda_1 = 0 (ablation) | Swift4D Lite · sear steak and flame steak | PSNR: 33.83 -> 33.31 | Hurt |
| use hash table size 2^15 and lambda_1 = 0 (ablation) | Swift4D Lite · sear steak and flame steak | training time: 25 -> 20 | Helped |
| vary the dynamic threshold zeta from 3 to 9 (ablation) | Swift4D · cook spinach | PSNR | No clear effect |
| use temporal importance pruning (ablation) | Swift4D · sear steak | visual artifact cleanliness | Helped |
Surprising findings
- Treating textureless regions of moving objects as static can improve quality by reducing the number of dynamic points, according to the authors' discussion.
Limitations
- Only supports multi-view scenes; does not support monocular datasets.
- Does not cover human reconstruction.
- Decomposition is incomplete: textureless dynamic regions may be labeled static and static regions near moving objects may be labeled dynamic.
Experimental setup
Datasets: N3DV MeetRoom Basketball court
Benchmarks: N3DV MeetRoom Basketball court
Metrics: PSNR DSSIM LPIPS training time FPS model size
Baselines: DyNeRF NeRFPlayer HexPlane K-Planes 4DGS 3DGStream SpaceTimeGS Real-Time4DGS StreamRF Plenoxel I-NGP 3DGS
Builds on: 3D Gaussian Splatting Instant-NGP 4DGS 3DGStream
Closest papers by meaning
- Grid4D: 4D Decomposed Hash Encoding for High-Fidelity Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.92)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · 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.