Research map / 3D vision / Gaussian splatting
Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes
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 DG-Mesh, which jointly optimizes deformable 3D Gaussians and a differentiable Poisson/marching-cubes mesh to reconstruct time-consistent, high-fidelity dynamic meshes with cross-frame vertex correspondence.
Problem
Obtaining high-quality, detailed, and time-consistent meshes with cross-frame vertex correspondence from dynamic monocular or multi-view observations is difficult; prior volumetric and template-based methods fail on thin structures, topology changes, or produce oversized meshes.
Approach
Maintain canonical 3D Gaussians deformed per timestep; convert deformed Gaussians to meshes via differentiable Poisson solver (DPSR) and differentiable marching cubes; apply Gaussian-Mesh Anchoring (mesh-guided densification/pruning so Gaussians map 1-1 to faces) and a cycle-consistent forward/backward deformation loss to propagate anchored Gaussians back to canonical space, yielding mesh-to-mesh
Key results
- On the DG-Mesh synthetic dataset, DG-Mesh achieves lower Chamfer Distance and EMD than all baselines on most scenes (e.g., Bird CD 0.510 vs 0.742 for next best K-Plane) and higher mesh-rendering PSNR (e.g., Horse 30.865 vs 22.395).
- On D-NeRF, DG-Mesh reports mesh-rendering PSNR of about 30 dB on average, e.g., Bouncingballs 29.145 and Jumpingjacks 31.769 versus baseline volumetric methods.
- On Nerfies, DG-Mesh achieves 10.9 mesh FPS versus 0.10 for Deformable-GS, with mesh rendering PSNR 27.238 vs 28.310 for Deformable-GS image PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DG-Mesh instead of the strongest listed baseline | dynamic mesh reconstruction · Duck, Horse, Bird, Beagle, Torus2sphere, and Girlwalk | CD | Helped |
| use DG-Mesh instead of the strongest listed baseline | D-NeRF mesh rendering · Bouncingballs and Jumpingjacks | mesh PSNR | Helped |
| use DG-Mesh instead of the strongest listed baseline | D-NeRF mesh rendering · Lego | mesh PSNR | Mixed |
| add Laplacian regularization with weight 1000 rather than no Laplacian regularization (ablation) | DG-Mesh · mesh reconstruction and rendering | mesh PSNR: 31.947 -> 33.890 | Helped |
| increase Laplacian weight from 1000 to 2000 (ablation) | DG-Mesh · mesh reconstruction and rendering | mesh PSNR: 33.890 -> 32.891 | Hurt |
| add Gaussian-Mesh Anchoring (ablation) | DG-Mesh · mesh reconstruction | CD: 0.685 -> 0.516 | Helped |
| add Gaussian-Mesh Anchoring (ablation) | DG-Mesh · mesh reconstruction | EMD: 0.146 -> 0.128 | Helped |
| add Gaussian-Mesh Anchoring (ablation) | DG-Mesh · mesh reconstruction | mesh PSNR: 31.283 -> 32.734 | Helped |
| change anchoring interval from 50 to 100 (ablation) | DG-Mesh · mesh reconstruction | CD: 0.314 -> 0.306 | Helped |
| increase anchoring interval from 100 to 500 (ablation) | DG-Mesh · mesh reconstruction | CD: 0.306 -> 0.320 | Hurt |
| fine-tune baselines with mesh-rendering loss (ablation) | K-Plane, TiNeuVox, and Deformable-GS versus DG-Mesh · fine-tuned mesh rendering | mesh PSNR: 24.683; 24.149; 22.519 -> 29.782 | Helped |
Surprising findings
- Fine-tuning baselines with a mesh rendering loss still leaves DG-Mesh ahead in mesh rendering PSNR, though the baselines' geometry quality remains worse.
Limitations
- Applied mainly to foreground objects, requiring segmentation masks for real videos.
- Correspondence becomes difficult under significant topology changes, such as new objects or parts appearing over time.
- Like other dynamic Gaussian methods, struggles with extreme object motion or imperfect segmentation.
Experimental setup
Datasets: D-NeRF DG-Mesh dataset (rendered synthetic) Nerfies DyCheck Unbiased4D NeuralActor self-captured iPhone videos
Benchmarks: DG-Mesh dataset D-NeRF Nerfies
Metrics: Chamfer Distance Earth Mover Distance PSNR mesh rendering PSNR SSIM LPIPS FPS
Baselines: D-NeRF K-Planes HexPlane TiNeuVox 4D-GS Deformable-GS 4DGS SC-GS
Builds on: 3D Gaussian Splatting Deformable 3D Gaussians Differentiable Poisson Surface Reconstruction Differentiable Marching Cubes Nvdiffrast
Closest papers by meaning
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering (CVPR 2024 · similarity 0.90)
- D-MiSo: Editing Dynamic 3D Scenes using Multi-Gaussians Soup (NeurIPS 2024 · similarity 0.89)
- FreeTimeGS: Free Gaussian Primitives at Anytime Anywhere for Dynamic Scene Reconstruction (CVPR 2025 · similarity 0.89)
- Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle (CVPR 2024 · similarity 0.89)
- Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling (AAAI 2025 · 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.