Reproduce “Physics-Grounded Motion Forecasting via Equation Discovery for Trajectory-Guided Image-to-Video Generation” 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
Tao Feng, Xianbing Zhao, Zhenhua Chen, Tien Tsin Wong, Hamid Rezatofighi, Gholamreza Haffari, Lizhen Qu
A neuro-symbolic framework merges equation discovery with trajectory-guided video generation, enabling the creation of future video sequences where object motion adheres to learned physical laws. The method, using retrieval-based symbolic regression (ReSR), improves both equation discovery efficiency and the physical consistency and visual quality of generated videos, outperforming data-driven baselines.
Key insights
- Explicitly embedding physics knowledge via discovered symbolic equations significantly enhances the physical plausibility and consistency of object motion within generated videos.
- A retrieval-based pre-training mechanism for symbolic regression effectively improves search efficiency, convergence speed, and accuracy by providing informed initial candidates from an equation bank.
- While trajectory guidance improves global motion consistency, current data-driven generative models may still struggle with fine-grained physical details, indicating a need for stronger physical inductive biases beyond high-level conditioning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.