Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation

Jinyang Yuan, Bin Li, Xiangyang Xue
Proceedings of the 36th International Conference on Machine Learning, PMLR 97:7222-7231, 2019.

Abstract

We present a deep generative model which explicitly models object occlusions for compositional scene representation. Latent representations of objects are disentangled into location, size, shape, and appearance, and the visual scene can be generated compositionally by integrating these representations and an infinite-dimensional binary vector indicating presences of objects in the scene. By training the model to learn spatial dependences of pixels in the unsupervised setting, the number of objects, pixel-level segregation of objects, and presences of objects in overlapping regions can be estimated through inference of latent variables. Extensive experiments conducted on a series of specially designed datasets demonstrate that the proposed method outperforms two state-of-the-art methods when object occlusions exist.

Cite this Paper

BibTeX
@InProceedings{pmlr-v97-yuan19b, title = {Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation}, author = {Yuan, Jinyang and Li, Bin and Xue, Xiangyang}, booktitle = {Proceedings of the 36th International Conference on Machine Learning}, pages = {7222--7231}, year = {2019}, editor = {Chaudhuri, Kamalika and Salakhutdinov, Ruslan}, volume = {97}, series = {Proceedings of Machine Learning Research}, month = {09--15 Jun}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v97/yuan19b/yuan19b.pdf}, url = {https://proceedings.mlr.press/v97/yuan19b.html}, abstract = {We present a deep generative model which explicitly models object occlusions for compositional scene representation. Latent representations of objects are disentangled into location, size, shape, and appearance, and the visual scene can be generated compositionally by integrating these representations and an infinite-dimensional binary vector indicating presences of objects in the scene. By training the model to learn spatial dependences of pixels in the unsupervised setting, the number of objects, pixel-level segregation of objects, and presences of objects in overlapping regions can be estimated through inference of latent variables. Extensive experiments conducted on a series of specially designed datasets demonstrate that the proposed method outperforms two state-of-the-art methods when object occlusions exist.} }
Endnote
%0 Conference Paper %T Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation %A Jinyang Yuan %A Bin Li %A Xiangyang Xue %B Proceedings of the 36th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2019 %E Kamalika Chaudhuri %E Ruslan Salakhutdinov %F pmlr-v97-yuan19b %I PMLR %P 7222--7231 %U https://proceedings.mlr.press/v97/yuan19b.html %V 97 %X We present a deep generative model which explicitly models object occlusions for compositional scene representation. Latent representations of objects are disentangled into location, size, shape, and appearance, and the visual scene can be generated compositionally by integrating these representations and an infinite-dimensional binary vector indicating presences of objects in the scene. By training the model to learn spatial dependences of pixels in the unsupervised setting, the number of objects, pixel-level segregation of objects, and presences of objects in overlapping regions can be estimated through inference of latent variables. Extensive experiments conducted on a series of specially designed datasets demonstrate that the proposed method outperforms two state-of-the-art methods when object occlusions exist.
APA
Yuan, J., Li, B. & Xue, X.. (2019). Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation. Proceedings of the 36th International Conference on Machine Learning, in Proceedings of Machine Learning Research 97:7222-7231 Available from https://proceedings.mlr.press/v97/yuan19b.html.

Related Material