DeNF: Unsupervised Scene-Decompositional Normalizing Flows

Farhad Ghazvinian Zanjani, Hanno Ackermann, Daniel Dijkman, Fatih Porikli

Abstract: Unsupervised object-centric scene decomposition models can learn compositional and hierarchical representations of multi-object scene data that allow the abstraction of the data into object entities and spaces. However, previous approaches, either based on VAE or GAN frameworks, have no guarantee of preserving particular aspects of the image in scene reconstruction. In this work, we propose the first probabilistic model called DeNF. Based on recent advances in normalizing flows, we represent the scene as a mixture of bidirectional flows that map a set of structured prior distributions into the scene data distribution. The bijective mapping of DeNF yields an efficient sampling and density evaluation in training time. Furthermore, it improves the fidelity of the scene's visual contents in the reconstruction process. In our experiments on real and synthetic image data for unsupervised scene decomposition, DeNF achieves competitive results.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning