Research map / 3D vision / Depth and stereo
MonoPhysics: Estimating Geometry, Appearance, and Physical Parameters from Monocular Videos
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 MonoPhysics, a monocular inverse-physics framework that jointly optimizes 3DGS geometry, appearance, and MPM material parameters using scale reparameterization, physics-aware geometry refinement, and a differentiable position map.
Problem
Recovering physical parameters and future dynamics of deformable objects from a single-camera video, where multi-view constraints that resolve scale and 3D geometry are absent.
Approach
Operates in camera space with a single learnable global scale scalar, makes per-particle volumes adaptive via iterative partition-of-unity, uses physical importance in particle relocation, and adds a position-map reparameterization that gives pixel-coordinate gradients to Gaussians, all coupled through a differentiable MPM simulator on 3DGS particles.
Key results
- On Vid2Sim, monocular MonoPhysics achieves MAE log E of 0.45, on par with multi-view Vid2Sim with 12 cameras (0.51).
- On the synthetic elastic set, MonoPhysics reduces CD to 24 vs 421 for GIC and gets PSNR 21.1 vs 18.5 for GIC.
- Ablation on plasticine subset: combining flow and silhouette losses yields CD 19 versus 64 without silhouette/flow (distribution regularizer only).
Surprising findings
- Optical flow supervision alone does not reduce Chamfer distance in the ablation.
- On the synthetic plasticine set, yield stress is the only plasticine parameter where any method beats the Oracle Midpoint baseline.
Limitations
- Assumes known camera parameters, known ground plane, and known constitutive model, limiting in-the-wild use.
- All evaluated sequences are gravity-driven falls onto a flat plane; other scenarios untested.
- Geometry refinement reduces but does not eliminate geometric errors in unobserved regions.
Experimental setup
Datasets: Vid2Sim Google Scanned Objects SpringGaus real captures
Benchmarks: Vid2Sim MonoPhysics synthetic dataset SpringGaus real captures
Metrics: PSNR SSIM LPIPS Chamfer Distance Earth Mover's Distance MAE log E MAE Poisson's ratio MAE log sigma_y MAE initial velocity
Baselines: PAC-NeRF SpringGaus GIC Vid2Sim Oracle Midpoint
Builds on: PAC-NeRF SpringGaus GIC Vid2Sim 3DGS-MCMC
Closest papers by meaning
- Vid2Sim: Generalizable, Video-based Reconstruction of Appearance, Geometry and Physics for Mesh-free Simulation (CVPR 2025 · similarity 0.88)
- OmniPhysGS: 3D Constitutive Gaussians for General Physics-Based Dynamics Generation (ICLR 2025 · similarity 0.88)
- MonoSplat: Generalizable 3D Gaussian Splatting from Monocular Depth Foundation Models (CVPR 2025 · similarity 0.87)
- World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.87)
- UniK3D: Universal Camera Monocular 3D Estimation (CVPR 2025 · similarity 0.86)
- PhysGen3D: Crafting a Miniature Interactive World from a Single Image (CVPR 2025 · 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.