The authors introduce the task of Scene Graph Anticipation (SGA) to forecast future interactions between objects using a novel approach called SceneSayer, leveraging object-centric representations and continuous-time dynamics modeling. The main thesis of the author is to propose a method, SceneSayer, that anticipates future pair-wise relationships between objects by leveraging object-centric representations and continuous-time dynamics modeling.
SceneSayer models future object relationships using NeuralODE and NeuralSDE.