IntPhys: A Framework and Benchmark for Visual Intuitive Physics Reasoning
In order to reach human performance on complexvisual tasks, artificial systems need to incorporate a sig-nificant amount of understanding of the world in termsof macroscopic objects, movements, forces, etc. Inspiredby work on intuitive physics in infants, we propose anevaluation benchmark which diagnoses how much a givensystem understands about physics by testing whether itcan tell apart well matched videos of possible versusimpossible events constructed with a game engine. Thetest requires systems to compute a physical plausibilityscore over an entire video. It is free of bias and cantest a range of basic physical reasoning concepts. Wethen describe two Deep Neural Networks systems aimedat learning intuitive physics in an unsupervised way,using only physically possible videos. The systems aretrained with a future semantic mask prediction objectiveand tested on the possible versus impossible discrimi-nation task. The analysis of their results compared tohuman data gives novel insights in the potentials andlimitations of next frame prediction architectures.