Sub-GAN: An Unsupervised Generative Model via Subspaces

Jie Liang, Jufeng Yang, Hsin Ying Lee, Kai Wang, Ming Hsuan Yang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

The recent years have witnessed significant growth in constructing robust generative models to capture informative distributions of natural data. However, it is difficult to fully exploit the distribution of complex data, like images and videos, due to the high dimensionality of ambient space. Sequentially, how to effectively guide the training of generative models is a crucial issue. In this paper, we present a subspace-based generative adversarial network (Sub-GAN) which simultaneously disentangles multiple latent subspaces and generates diverse samples correspondingly. Since the high-dimensional natural data usually lies on a union of low-dimensional subspaces which contain semantically extensive structure, Sub-GAN incorporates a novel clusterer that can interact with the generator and discriminator via subspace information. Unlike the traditional generative models, the proposed Sub-GAN can control the diversity of the generated samples via the multiplicity of the learned subspaces. Moreover, the Sub-GAN follows an unsupervised fashion to explore not only the visual classes but the latent continuous attributes. We demonstrate that our model can discover meaningful visual attributes which is hard to be annotated via strong supervision, e.g., the writing style of digits, thus avoid the mode collapse problem. Extensive experimental results show the competitive performance of the proposed method for both generating diverse images with satisfied quality and discovering discriminative latent subspaces.

Original languageEnglish
Title of host publicationComputer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings
EditorsVittorio Ferrari, Cristian Sminchisescu, Yair Weiss, Martial Hebert
PublisherSpringer Verlag
Pages726-743
Number of pages18
ISBN (Print)9783030012519
DOIs
Publication statusPublished - 2018 Jan 1
Event15th European Conference on Computer Vision, ECCV 2018 - Munich, Germany
Duration: 2018 Sep 82018 Sep 14

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11215 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other15th European Conference on Computer Vision, ECCV 2018
CountryGermany
CityMunich
Period18/9/818/9/14

Fingerprint

Generative Models
Subspace
Discriminators
Attribute
Digit
Dimensionality
Multiplicity
Union
High-dimensional
Generator
Experimental Results

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Cite this

Liang, J., Yang, J., Lee, H. Y., Wang, K., & Yang, M. H. (2018). Sub-GAN: An Unsupervised Generative Model via Subspaces. In V. Ferrari, C. Sminchisescu, Y. Weiss, & M. Hebert (Eds.), Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings (pp. 726-743). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11215 LNCS). Springer Verlag. https://doi.org/10.1007/978-3-030-01252-6_43
Liang, Jie ; Yang, Jufeng ; Lee, Hsin Ying ; Wang, Kai ; Yang, Ming Hsuan. / Sub-GAN : An Unsupervised Generative Model via Subspaces. Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings. editor / Vittorio Ferrari ; Cristian Sminchisescu ; Yair Weiss ; Martial Hebert. Springer Verlag, 2018. pp. 726-743 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{8e1bddc2a5684e9192d0a07c19b8ffe6,
title = "Sub-GAN: An Unsupervised Generative Model via Subspaces",
abstract = "The recent years have witnessed significant growth in constructing robust generative models to capture informative distributions of natural data. However, it is difficult to fully exploit the distribution of complex data, like images and videos, due to the high dimensionality of ambient space. Sequentially, how to effectively guide the training of generative models is a crucial issue. In this paper, we present a subspace-based generative adversarial network (Sub-GAN) which simultaneously disentangles multiple latent subspaces and generates diverse samples correspondingly. Since the high-dimensional natural data usually lies on a union of low-dimensional subspaces which contain semantically extensive structure, Sub-GAN incorporates a novel clusterer that can interact with the generator and discriminator via subspace information. Unlike the traditional generative models, the proposed Sub-GAN can control the diversity of the generated samples via the multiplicity of the learned subspaces. Moreover, the Sub-GAN follows an unsupervised fashion to explore not only the visual classes but the latent continuous attributes. We demonstrate that our model can discover meaningful visual attributes which is hard to be annotated via strong supervision, e.g., the writing style of digits, thus avoid the mode collapse problem. Extensive experimental results show the competitive performance of the proposed method for both generating diverse images with satisfied quality and discovering discriminative latent subspaces.",
author = "Jie Liang and Jufeng Yang and Lee, {Hsin Ying} and Kai Wang and Yang, {Ming Hsuan}",
year = "2018",
month = "1",
day = "1",
doi = "10.1007/978-3-030-01252-6_43",
language = "English",
isbn = "9783030012519",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Verlag",
pages = "726--743",
editor = "Vittorio Ferrari and Cristian Sminchisescu and Yair Weiss and Martial Hebert",
booktitle = "Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings",
address = "Germany",

}

Liang, J, Yang, J, Lee, HY, Wang, K & Yang, MH 2018, Sub-GAN: An Unsupervised Generative Model via Subspaces. in V Ferrari, C Sminchisescu, Y Weiss & M Hebert (eds), Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11215 LNCS, Springer Verlag, pp. 726-743, 15th European Conference on Computer Vision, ECCV 2018, Munich, Germany, 18/9/8. https://doi.org/10.1007/978-3-030-01252-6_43

Sub-GAN : An Unsupervised Generative Model via Subspaces. / Liang, Jie; Yang, Jufeng; Lee, Hsin Ying; Wang, Kai; Yang, Ming Hsuan.

Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings. ed. / Vittorio Ferrari; Cristian Sminchisescu; Yair Weiss; Martial Hebert. Springer Verlag, 2018. p. 726-743 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11215 LNCS).

Research output: Chapter in Book/Report/Conference proceedingConference contribution

TY - GEN

T1 - Sub-GAN

T2 - An Unsupervised Generative Model via Subspaces

AU - Liang, Jie

AU - Yang, Jufeng

AU - Lee, Hsin Ying

AU - Wang, Kai

AU - Yang, Ming Hsuan

PY - 2018/1/1

Y1 - 2018/1/1

N2 - The recent years have witnessed significant growth in constructing robust generative models to capture informative distributions of natural data. However, it is difficult to fully exploit the distribution of complex data, like images and videos, due to the high dimensionality of ambient space. Sequentially, how to effectively guide the training of generative models is a crucial issue. In this paper, we present a subspace-based generative adversarial network (Sub-GAN) which simultaneously disentangles multiple latent subspaces and generates diverse samples correspondingly. Since the high-dimensional natural data usually lies on a union of low-dimensional subspaces which contain semantically extensive structure, Sub-GAN incorporates a novel clusterer that can interact with the generator and discriminator via subspace information. Unlike the traditional generative models, the proposed Sub-GAN can control the diversity of the generated samples via the multiplicity of the learned subspaces. Moreover, the Sub-GAN follows an unsupervised fashion to explore not only the visual classes but the latent continuous attributes. We demonstrate that our model can discover meaningful visual attributes which is hard to be annotated via strong supervision, e.g., the writing style of digits, thus avoid the mode collapse problem. Extensive experimental results show the competitive performance of the proposed method for both generating diverse images with satisfied quality and discovering discriminative latent subspaces.

AB - The recent years have witnessed significant growth in constructing robust generative models to capture informative distributions of natural data. However, it is difficult to fully exploit the distribution of complex data, like images and videos, due to the high dimensionality of ambient space. Sequentially, how to effectively guide the training of generative models is a crucial issue. In this paper, we present a subspace-based generative adversarial network (Sub-GAN) which simultaneously disentangles multiple latent subspaces and generates diverse samples correspondingly. Since the high-dimensional natural data usually lies on a union of low-dimensional subspaces which contain semantically extensive structure, Sub-GAN incorporates a novel clusterer that can interact with the generator and discriminator via subspace information. Unlike the traditional generative models, the proposed Sub-GAN can control the diversity of the generated samples via the multiplicity of the learned subspaces. Moreover, the Sub-GAN follows an unsupervised fashion to explore not only the visual classes but the latent continuous attributes. We demonstrate that our model can discover meaningful visual attributes which is hard to be annotated via strong supervision, e.g., the writing style of digits, thus avoid the mode collapse problem. Extensive experimental results show the competitive performance of the proposed method for both generating diverse images with satisfied quality and discovering discriminative latent subspaces.

UR - http://www.scopus.com/inward/record.url?scp=85055128592&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85055128592&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-01252-6_43

DO - 10.1007/978-3-030-01252-6_43

M3 - Conference contribution

AN - SCOPUS:85055128592

SN - 9783030012519

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 726

EP - 743

BT - Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings

A2 - Ferrari, Vittorio

A2 - Sminchisescu, Cristian

A2 - Weiss, Yair

A2 - Hebert, Martial

PB - Springer Verlag

ER -

Liang J, Yang J, Lee HY, Wang K, Yang MH. Sub-GAN: An Unsupervised Generative Model via Subspaces. In Ferrari V, Sminchisescu C, Weiss Y, Hebert M, editors, Computer Vision – ECCV 2018 - 15th European Conference, 2018, Proceedings. Springer Verlag. 2018. p. 726-743. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-030-01252-6_43