Long Term Boundary Extrapolation for Deterministic Motion
Apratim Bhattacharyya, Mateusz Malinowski, Mario Fritz · Max Planck Digital Library · 2016
We propose models for long-term prediction of boundaries that are learned from observations without making any strong model assumptions on objects or scene.We evaluate our approach in a billiard scenario that is only governed by the Newtonian laws of physics and is therefore fully deterministic.We argue that any model that succeeds in this task must have derived some notion of "intuitive physics".Our Recursive Convolutional Multi-Scale architecture turns out to be most effective.