Unconditional Scene Graph Generation (ICCV 2021)

Sarthak Garg *     Helisa Dhamo *     Azade Farshad      Sabrina Musatian      Nassir Navab     Federico Tombari    

Technical University of Munich    Google

* The first two authors contributed equally.




Scene graphs, composed of nodes as objects and directed-edges as relationships among objects, offer an alternative representation of a scene that is more semantically grounded than images. We hypothesize that a generative model for scene graphs might be able to learn the underlying semantic structure of real-world scenes more effectively than images, and hence, generate realistic novel scenes in the form of scene graphs. In this work, we explore a new task for the unconditional generation of semantic scene graphs using a generative approach. For this task, a deep auto-regressive model called SceneGraphGen is developed, which can directly learn the probability distribution over labelled and directed graphs using a hierarchical recurrent architecture. We demonstrate the application of the generated graphs in image synthesis, anomaly detection and scene graph completion.

Paper

International Conference on Computer Vision (ICCV) 2021
Paper PDF | arXiv
@inproceedings{scenegraphgen2021,
  title={Unconditional Scene Graph Generation},
  author={Garg, Sarthak and Dhamo, Helisa and Farshad, Azade and Musatian, Sabrina and Navab, Nassir and Tombari, Federico},
  booktitle={IEEE International Conference on Computer Vision (ICCV)},
  year={2021}
}
  

Downloads

The cleaned scene graph dataset used in our paper, based on Visual Genome, can be downloaded here.

Source Code

The source code for this work can be found here.

Contact

For questions regarding the method or the source code, please contact sarthak.garg[at]tum.de.