Scene Graph Prediction

Scene graph prediction aims to computationally represent a scene by identifying objects and their relationships, creating a structured graph. Current research focuses on improving accuracy and generalizability, particularly through incorporating hierarchical relationships, commonsense knowledge, and language models to handle open-vocabulary objects and relationships, often leveraging techniques like message passing and contrastive learning within various model architectures. This work is significant for advancing scene understanding in both 2D and 3D environments, with applications ranging from robotics and autonomous systems to improved image and video analysis.