Reproduce “PhysConvex: Physics-Informed 3D Dynamic Convex Radiance Fields for Reconstruction and Simulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dan Wang, Xinrui Cui, Serge Belongie, Ravi Ramamoorthi
PhysConvex reconstructs object appearance, geometry, and physical properties from video. It represents objects as deformable convex shapes whose boundaries change locally, and simulates them with compact physics-informed deformation modes. On 12 GSO test objects, it reports PSNR 30.00 versus 28.57 for Vid2Sim trained without LGM initialization. Future prediction improves PSNR and SSIM, but not FoVVDP, compared with Vid2Sim.
Key insights
- A novel boundary-driven dynamic convex representation can effectively bridge the gap between visually appealing neural representations and physically accurate simulation, accommodating flexible, spatially adaptive deformation.
- Integrating neural skinning eigenmodes as low-dimensional, material- and shape-aware deformation bases within a reduced-order simulation framework allows for efficient and robust modeling of complex elastodynamics.
- An end-to-end differentiable system enables joint optimization of appearance, geometry, and underlying physical properties (e.g., Young’s modulus, Poisson’s ratio) directly from multi-view video observations, enhancing coherence and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.