Research map / 3D vision / Gaussian splatting
ReMatching Dynamic Reconstruction Flow
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 ReMatching, a velocity-field-based prior framework with a flow-matching-style loss that projects a simulation-free time-dependent Gaussian Splatting reconstruction onto prior classes (rigid, divergence-free, piece-wise) to improve novel-timestamp dynamic reconstruction.
Problem
Novel-view dynamic scene reconstruction from sparse multi-view images struggles to generalize to unseen viewpoints and timestamps; data-driven or physical priors often reduce fidelity to the input.
Approach
Formulate the projection of a simulation-free time-dependent reconstruction function onto a velocity-field prior class as a flow-matching problem (ReMatching loss), solved in closed form via linear least squares over linear parameterizations (rigid, divergence-free basis, and adaptive piece-wise combinations), and add it as a weighted regularizer to the reconstruction loss.
Key results
- On D-NeRF unseen frames, the P3 variant of ReMatching improves over D3G on Bouncing Balls (PSNR 41.84 vs 41.52) and JumpingJacks (37.72 vs 37.59), with P4/P5 reaching 38.01 on JumpingJacks.
- On HyperNeRF, the P4 variant reaches PSNR 32.67 on Split Cookie vs 32.61 for D3G, and LPIPS 0.0937 vs 0.0971.
- Applied on top of GA3D, the framework improves PSNR on D-NeRF T-Rex from 37.70 to 39.18 for the P4/P5 variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the divergence-free prior P3 instead of the D3G baseline | D-NeRF reconstruction · D-NeRF Bouncing Balls | PSNR: 41.52 -> 41.84 | Helped |
| Use adaptive-combination prior P4/P5 instead of P3 | ReMatching · D-NeRF Hell Warrior | PSNR: 41.59 -> 41.69 | Helped |
| Use adaptive-combination prior P4/P5 instead of P3 | ReMatching · D-NeRF Lego | PSNR: 24.89 -> 24.95 | No clear effect |
| Use adaptive-combination prior P4/P5 instead of P3 | ReMatching · D-NeRF Hook | PSNR: 37.19 -> 37.39 | Helped |
| Apply the divergence-free prior P3 instead of D3G | ReMatching · D-NeRF Mutant | PSNR: 42.09 -> 42.13 | Hurt |
| Apply adaptive-combination prior P4/P5 instead of D3G | ReMatching · D-NeRF Stand Up | PSNR: 43.85 -> 44.31 | Helped |
| Apply divergence-free prior P3 instead of D3G | ReMatching · HyperNeRF | PSNR | Mixed |
| Apply adaptive-combination prior P4 instead of D3G | ReMatching · HyperNeRF | PSNR | Mixed |
| Add ReMatching P4/P5 to the GA3D backbone | GA3D · D-NeRF Bouncing Balls | PSNR: 43.42 -> 43.62 | Helped |
Surprising findings
- On HyperNeRF, the P3 variant underperforms D3G on several scenes (e.g., Lemon PSNR 27.91 vs 28.65 and Split Cookie 31.31 vs 32.61), indicating that the prior does not uniformly help.
Limitations
- Gains over baselines are small in most scenes and some scenes show regressions.
- Priors are restricted to linear parameterizations (rigid, divergence-free, piece-wise) that may be too simple for complex real-world physics such as liquids and gases.
- Requires the reconstruction model to be simulation-free and differentiable in time and space, limiting applicability.
Experimental setup
Datasets: D-NeRF HyperNeRF Dynamic Scenes dataset
Benchmarks: D-NeRF HyperNeRF Dynamic Scenes dataset
Metrics: PSNR SSIM LPIPS
Baselines: D3G GA3D NPG K-Planes
Builds on: Deformable 3D Gaussians 3D Geometry-aware Deformable Gaussians alternating projections method Gaussian Splatting
Closest papers by meaning
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.87)
- Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 2025 · similarity 0.87)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.87)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.86)
- HAIF-GS: Hierarchical and Induced Flow-Guided Gaussian Splatting for Dynamic Scene (NeurIPS 2025 · similarity 0.86)
- ReconFusion: 3D Reconstruction with Diffusion Priors (CVPR 2024 · similarity 0.86)
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.