Research map / 3D vision / Gaussian splatting
DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions
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 DySurface, which couples a deformable 3D Gaussian branch with a voxel-conditioned dynamic SDF branch and a mesh refinement stage to extract temporally consistent 4D meshes from video.
Problem
Reconstructing temporally consistent, watertight geometric surfaces of dynamic scenes, where photometric-only 3DGS/NeRF training yields geometric ambiguity, fragmented surfaces, and inaccurate meshes.
Approach
Bridges forward-deformed 3DGS and backward-warped volumetric SDF rendering by voxelizing deformed Gaussians to condition a backward deformation MLP and SDF field, with cycle-consistency and SDF-to-Gaussian anchoring losses, then refines a canonical mesh with the forward Gaussian motion field.
Key results
- On D-NeRF averages, DySurface reaches vIoU 0.3932 and CD 0.0102, versus 0.3814/0.0137 for DG-Mesh and 0.3426/0.0111 for D-2DGS, while PSNR is 31.09 vs 32.24 for GaGS.
- On the DG-Mesh dataset, DySurface attains CD 0.0078 (best) and PSNR 30.92, but vIoU 0.7571 trails DG-Mesh's 0.7602.
- Temporal metrics on DG-Mesh: TNAE drops from 15.08 (DG-Mesh) to 11.30 and trajectory position error from 0.00637 to 0.00508.
Surprising findings
- On D-NeRF, the w/o VoxGS-DSDF ablation has higher vIoU than some baselines but the full model's vIoU is only marginally above DG-Mesh on several time steps, and the 2DGS variant gives higher vIoU on HellWarrior.
- GaGS achieves higher PSNR than DySurface on D-NeRF (32.24 vs 31.09), so the geometric gains come with a small photometric cost.
Limitations
- Struggles with extreme topological changes such as tearing, fracturing, or merging, since a single canonical space is used.
- Minor vertex noise and normal shifts in the mesh refinement stage; stronger Laplacian smoothing reduces noise but can remove fine detail.
- Unconstrained in-the-wild videos with severe single-view occlusions remain challenging.
Experimental setup
Datasets: D-NeRF DG-Mesh Nerfies DyCheck
Benchmarks: D-NeRF DG-Mesh Nerfies DyCheck
Metrics: vIoU Chamfer Distance PSNR SSIM LPIPS EPE3D TNAE Trajectory Position Error
Baselines: TiNeuVox NeuS2 4DGS GaGS DG-Mesh D-2DGS 3DGS-to-PC GS2Mesh
Builds on: GaGS NeuS 3D Gaussian Splatting DG-Mesh Neuralangelo
Closest papers by meaning
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.90)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.88)
- Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes (ICLR 2025 · similarity 0.88)
- GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes (NeurIPS 2024 · similarity 0.88)
- MAtCha Gaussians: Atlas of Charts for High-Quality Geometry and Photorealism From Sparse Views (CVPR 2025 · similarity 0.88)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.88)
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.