/Recurrent-Interaction-Network

Pytorch implementation of Recurrent Interaction Networks, from "Occlusion resistant learning of intuitive physics from videos"

Primary LanguagePythonMIT LicenseMIT

Recurrent-Interaction-Network

Pytorch implementation of (Recurrent) Interaction Network, from "Occlusion resistant learning of intuitive physics from videos"

The code is divided in two modules:

  • rin: Recurrent Interaction Network learning to predict future trajectories of objects
  • cnr: Compositional Neural Network learning to reconstruct a segmentation mask given the position of objects in the scene

datasets/ gathers datasets used for experiments.