Research map / 3D vision / Gaussian splatting
3D Geometry-Aware Deformable Gaussian Splatting for Dynamic View Synthesis
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 a 3D geometry-aware deformable Gaussian Splatting method that voxelizes canonical Gaussians with a sparse 3D U-Net to extract local geometric features that drive a deformation MLP for dynamic view synthesis from monocular video.
Problem
Dynamic view synthesis from monocular video requires geometrically coherent deformation modeling, which implicit NeRF-based deformation fields and point-wise or interpolated features fail to capture.
Approach
Treats canonical 3D Gaussians as a point cloud, extracts local geometric features with a sparse 3D U-Net over voxels, fuses them with point features, and predicts per-Gaussian position, rotation (continuous 6D) and scale deformations over time; density control is done on transformed Gaussians at each timestamp.
Key results
- On the D-NeRF synthetic dataset, the method reaches 38.0134 average PSNR in ablation (full model) and clearly outperforms 4D-GS on most scenes, e.g., 42.2101 PSNR vs 38.2610 on Jumping Jacks-like scenes such as Stand Up.
- On HyperNeRF averages over Cut Lemon, Chicken, 3D Printer and Split Cookie, the method reports 27.52 PSNR, 0.80 SSIM and 0.25 LPIPS versus 26.98 PSNR for 4D-GS.
- Ablation shows removing geometric features drops average PSNR from 38.01 to 37.58 and using a quaternion instead of 6D rotation drops it to 37.88 (per-table numbers).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add geometry-aware features to the canonical field (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 37.5757 -> 38.0134 | Helped |
| Use continuous 6D rotation representation instead of quaternion (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 37.8750 -> 38.0134 | Helped |
| Control density using transformed Gaussians at timestamps, not only canonical Gaussians (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 37.8026 -> 38.0134 | Helped |
| Use sparse-convolution geometry features instead of PointNet-like features (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 36.7353 -> 38.0134 | Helped |
| Use sparse-convolution geometry features instead of plane-projection features (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 35.9054 -> 38.0134 | Helped |
| Add the proposed components instead of using the vanilla model (ablation) | 3D Geometry-Aware Deformable Gaussian Splatting · Synthetic dataset ablation average | PSNR: 35.2307 -> 38.0134 | Helped |
| Use geometry-aware deformation instead of 4D-GS | Dynamic view synthesis on D-NeRF scenes · D-NeRF: T-Rex | PSNR: 33.3912 -> 39.0285 | Helped |
| Use geometry-aware deformation instead of 4D-GS | Dynamic view synthesis on HyperNeRF scenes · HyperNeRF average (Cut Lemon, Chicken, 3D Printer, Split Cookie) | PSNR: 26.98 -> 27.52 | Helped |
Surprising findings
- On HyperNeRF's Broom and Chicken-like scenes, the method does not beat all baselines (e.g., lower MS-SSIM than TiNeuVox on 3D Printer and Broom), and 3D-GS achieves much lower HyperNeRF scores than the NeRF-based methods.
Limitations
- Explicit motion modeling via foreground/background segmentation cues is not yet addressed.
- Rendering speed is modest (12 FPS), and performance on some real HyperNeRF scenes lags baselines.
Experimental setup
Datasets: D-NeRF HyperNeRF
Benchmarks: D-NeRF HyperNeRF
Metrics: PSNR SSIM LPIPS MS-SSIM
Baselines: 3D-GS D-NeRF TiNeuVox NDVG FDNeRF 4D-GS
Builds on: 3D Gaussian Splatting D-NeRF NeRF positional encoding sparse convolution 6D rotation representation
Closest papers by meaning
- LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors (CVPR 2025 · similarity 0.88)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.88)
- Template-free Articulated Gaussian Splatting for Real-time Reposable Dynamic View Synthesis (NeurIPS 2024 · similarity 0.88)
- SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes (CVPR 2024 · similarity 0.87)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.87)
- DisC-GS: Discontinuity-aware Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
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.