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 objectscnr: Compositional Neural Network learning to reconstruct a segmentation mask given the position of objects in the scene
datasets/ gathers datasets used for experiments.