Reproduce “PhysGaia: A Physics-Aware Benchmark with Multi-Body Interactions for Dynamic Novel View Synthesis” 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
Mijeong Kim, Gunhee Kim, Jungyoon Choi, Wonjae Roh, Bohyung Han
PhysGaia introduces a physics-aware benchmark dataset for Dynamic Novel View Synthesis (DyNVS), featuring 17 scenes with complex multi-body interactions across diverse materials like liquids, gases, and textiles, alongside ground-truth physical parameters and 3D trajectories. The benchmark reveals that current DyNVS methods struggle with physical realism and complex dynamics, achieving average PSNR scores below 30 and high trajectory distance (TD) and area under outlier percentages (AUOP) in dynamic scenes.
Key insights
- PhysGaia's high motion complexity and photorealism make it a robust benchmark, exhibiting a dynamic score of 0.444 and FID of 207.8, which is superior to existing physics-based datasets.
- Current state-of-the-art DyNVS methods (e.g., D-3DGS, 4DGS) achieve limited photorealism (PSNR below 30) and struggle significantly with physical plausibility in complex scenes on PhysGaia, showing high TD (e.g., D-3DGS: 4.01 in "box-smoke") and AUOP.
- Existing inverse physics methods (PAC-NeRF, GIC), designed for single-object scenarios, significantly underestimate material stiffness (e.g., Young's modulus) when applied to PhysGaia's multi-object interaction scenes, highlighting challenges in complex physics parameter estimation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.