Research map / 3D vision / Gaussian splatting
FreeTimeGS: Free Gaussian Primitives at Anytime Anywhere for Dynamic Scene Reconstruction
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 FreeTimeGS, a 4D Gaussian representation whose primitives appear at arbitrary times and locations with linear velocity motion and temporal opacity, plus a 4D opacity regularizer and periodic relocation, for real-time dynamic novel view synthesis of fast-motion scenes.
Problem
Reconstructing dynamic 3D scenes with complex, fast motion from multi-view video; deformation-field and canonical 3D Gaussian methods struggle to build long-range correspondences and optimize poorly under large motion.
Approach
Gaussian primitives are free in both space and time: each has a temporal Gaussian opacity (duration) and a linear velocity motion function, so it only needs short-range motion modeling; a stop-gradient opacity regularizer weighted by temporal opacity avoids high-opacity blocking of gradients, and periodic relocation reallocates low-opacity primitives.
Key results
- On Neural3DV, FreeTimeGS reaches 33.19 PSNR and 0.036 LPIPS, beating STGS (32.05 PSNR, 0.044 LPIPS).
- On SelfCap, it reaches 27.41 dB full-image PSNR and 29.38 dB dynamic-region PSNR at 467 FPS, versus 25.98/26.75 dB for 4DGS.
- On ENeRF-Outdoor, it achieves 25.36 PSNR and 454 FPS, versus 24.93 PSNR for STGS at 226 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FreeTimeGS instead of Ex4DGS | FreeTimeGS · Neural3DV | PSNR: 32.11 -> 33.19 | Helped |
| use FreeTimeGS instead of 4K4D | FreeTimeGS · ENeRF-Outdoor | PSNR: 25.28 -> 25.36 | Helped |
| use FreeTimeGS instead of 4DGS | FreeTimeGS · SelfCap (entire) | PSNR: 25.98 -> 27.41 | Helped |
| use FreeTimeGS instead of 4DGS | FreeTimeGS · SelfCap (dynamic) | PSNR: 26.75 -> 29.38 | Helped |
| use the proposed motion representation instead of 4DGS motion representation (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 28.10/26.92 -> 29.74/30.75 | Helped |
| add 4D regularization loss (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 28.68/29.09 -> 29.74/30.75 | Helped |
| add periodic relocation instead of using the 3DGS densify strategy (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 29.07/29.15 -> 29.74/30.75 | Helped |
| add proposed 4D velocity initialization instead of zero velocity initialization (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 28.33/27.06 -> 29.74/30.75 | Helped |
| set λreg to 0 instead of 1e−2 (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 29.74/30.75 -> 28.68/29.09 | Hurt |
| set λreg to 1e−3 instead of 1e−2 (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 29.74/30.75 -> 29.10/29.79 | Hurt |
| set λreg to 1e−1 instead of 1e−2 (ablation) | FreeTimeGS · SelfCap dance1 (entire/fastest) | PSNR: 29.74/30.75 -> 26.43/27.33 | Hurt |
Surprising findings
- Many primitives approach opacity 1, and that high opacity blocks gradient flow to other primitives, causing local minima on fast-moving regions.
Limitations
- Requires a lengthy per-scene reconstruction process; no optimization-free reconstruction.
- Does not support relighting since it lacks surface normal and material modeling.
- Regularization increases the number of Gaussians, requiring periodic relocation to mitigate.
Experimental setup
Datasets: Neural3DV ENeRF-Outdoor SelfCap
Benchmarks: Neural3DV ENeRF-Outdoor SelfCap
Metrics: PSNR DSSIM LPIPS FPS
Baselines: 4DGS STGS 4K4D ENeRF Ex4DGS HexPlane K-Planes MixVoxels Deformable-3DGS DyNeRF
Builds on: 3D Gaussian Splatting 4DGS Deformable 3D Gaussians ROMA
Closest papers by meaning
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.91)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.90)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.90)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · similarity 0.90)
- 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering (NeurIPS 2025 · similarity 0.90)
- Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation (AAAI 2025 · similarity 0.90)
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.