Extraction of Motion Change Points Based on the Physical Characteristics of Objects

Recently, it has become increasingly important for artificial intelligence to understand the real world through the ability of intuitive physics, which is our innate ability to understand the real world. Many previous studies aiming at real-world understanding have based image inference for real world recognition, usually based on inference from image features or recognition of objects in an image. In contrast, we propose a model to obtain and predict the change points of real-world motions represented by the potential hierarchical structure of the physical relations of the observed objects represented by graph embeddings. We conducted experiments on the CLEVRER dataset [1] to detect motion change points and extract predictive change points, and found that the proposed model is correctly trained as a predictive change point extraction model.