Scene Graph Generation

Scene graph generation identifies objects in images or videos and connects them with relationships, creating explicit representations that can support visual question answering, activity analysis, and robot planning. Common approaches combine object detectors with relation classifiers, graph neural networks, or transformers; newer work uses vision-language models, open-vocabulary learning, and diffusion to handle unfamiliar categories and structured outputs. The field spans image benchmarks such as Visual Genome, video datasets such as Action Genome, and 3D settings such as 3DSSG, with growing attention to temporal relations and embodied tasks. A central challenge is producing precise, useful relations despite long-tailed categories and incomplete annotations, while maintaining robust generalization and efficient inference.