Research map / 3D vision / Gaussian splatting
DN-4DGS: Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering
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 DN-4DGS, a two-stage deformation for canonical 3D Gaussians with a noise suppression strategy plus decoupled temporal-spatial feature aggregation, for real-time dynamic scene rendering.
Problem
Canonical 3D Gaussians for dynamic scenes contain noisy coordinates that propagate into the deformation field, and existing methods lack effective 4D (spatio-temporal) feature aggregation.
Approach
A two-stage deformation: a first deformation predicts coordinate offsets that re-distribute canonical Gaussian positions to suppress noise, then a second deformation applies KNN spatial aggregation on the denoised coordinates, while temporal aggregation over neighboring frames is done in the first stage.
Key results
- On PlenopticVideo, achieves 32.02 PSNR vs 31.15 for 4DGaussian, with LPIPS 0.043 and 112 FPS reported.
- On HyperNeRF with dense init, reaches 25.59 PSNR and 0.863 MS-SSIM vs 25.20 and 0.845 for 4DGaussian.
- On HyperNeRF with sparse init, improves PSNR by +1.51 over 4DGaussian (23.31 vs 21.80).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DN-4DGS noise suppression and temporal-spatial aggregation | 4DGaussian on PlenopticVideo · PlenopticVideo | PSNR: 31.15 -> 32.02 | Helped |
| Add DN-4DGS | 4DGaussian with sparse initialization · HyperNeRF Sparse Init | PSNR: 21.80 -> 23.31 (+1.51) | Helped |
| Add DN-4DGS | 4DGaussian with dense initialization · HyperNeRF Dense Init | PSNR: 25.20 -> 25.59 (+0.39) | Helped |
| Add DN-4DGS | NeRF-DS baselines · NeRF-DS scenes | PSNR | Mixed |
| Add learnable embedding Y (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.13 -> 33.51 | Helped |
| Set Y dimension to 4 instead of 16 (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.33 vs 33.51 | Hurt |
| Set Y dimension to 32 instead of 16 (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.40 vs 33.51 | Hurt |
| Set Y dimension to 64 instead of 16 (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.29 vs 33.51 | Hurt |
| Set temporal aggregation timestep to 0.5x instead of 1x (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.40 vs 33.51 | Hurt |
| Set temporal aggregation timestep to 2x instead of 1x (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.45 vs 33.51 | Hurt |
| Set spatial neighbor count K to 16 (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 33.39 (K=4), 33.51 (K=16), 33.35 (K=32) | Helped |
| Train the first deformation stage before the second, using 8000 first-stage rounds (ablation) | DN-4DGS · PlenopticVideo Flame Steak | PSNR: 32.89 (0 rounds), 33.51 (8000 rounds) | Helped |
Surprising findings
- Naively applying spatial aggregation on canonical Gaussians (single-stage) lowers quality, supporting the claim that noisy coordinates get amplified by KNN aggregation.
- Per-scene PlenopticVideo results show the method underperforms 4DGaussian on Coffee Martini and Flame Salmon in SSIM/PSNR for some scenes.
Limitations
- Two-stage training lacks simultaneous supervision, so second-stage coordinate deformation direction is uncontrollable.
- Slightly longer training time and storage than the 4DGaussian baseline due to the second deformation and learnable attribute Y.
- Offline training; unlike 3DGStream it cannot do online streaming reconstruction.
Experimental setup
Datasets: PlenopticVideo HyperNeRF NeRF-DS D-NeRF
Benchmarks: PlenopticVideo HyperNeRF NeRF-DS D-NeRF
Metrics: PSNR SSIM MS-SSIM LPIPS training time FPS storage
Baselines: DyNeRF NeRFPlayer HyperReel HexPlane K-Planes 4DGS E-D3DGS 4DGaussian D3DGS TiNeuVox Nerfies HyperNeRF
Builds on: 3D Gaussian Splatting 4DGaussians Deformable 3D Gaussians PointNet++ HexPlane
Closest papers by meaning
- 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering (NeurIPS 2025 · similarity 0.90)
- H3D-DGS: Exploring Heterogeneous 3D Motion Representation for Deformable 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle (CVPR 2024 · similarity 0.89)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.89)
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.