Physics-Informed Motion
Can physical laws improve video prediction when training data is scarce?
Research question
Can explicit dynamics improve video prediction when training data is small and object motion includes collisions or abrupt changes?
Method
- Detected object centers with OpenCV and fit quadratic trajectories within non-collision segments to derive velocity and acceleration.
- Used a working Pymunk simulator for collision-constrained motion, then trained a conditional GAN to render texture from predicted outlines and prior frames.
- Compared synchronized recorded, physics-based, and learned predictions across contact and no-contact scenes.
Evidence
The project page exposes all four evaluation clips, frame stepping, contact/no-contact ablations, three ConvLSTM seeds, the current paper, and a downloadable experiment package.
Boundary
The manuscript reports an exploratory grouped experiment over 20 distinct recordings with only two evaluation groups. It demonstrates a reproducible physics-informed prototype, not broad generalization.