TY - GEN

T1 - Learning disentangled representations via independent subspaces

AU - Awiszus, Maren

AU - Ackermann, Hanno

AU - Rosenhahn, Bodo

N1 - Funding information: The work is inspired by BIAS (”Bias and Discrimination in Big Data and Algorithmic Processing. Philosophical Assessments, Legal Dimensions, and Technical Solutions”), a project funded by the Volkswagen Foundation within the initiative ”AI and the Society of the Future” for which the last author is a Principal Investigator.

PY - 2019/10

Y1 - 2019/10

N2 - Image generating neural networks are mostly viewed as black boxes, where any change in the input can have a number of globally effective changes on the output. In this work, we propose a method for learning disentangled representations to allow for localized image manipulations. We use face images as our example of choice. Depending on the image region, identity and other facial attributes can be modified. The proposed network can transfer parts of a face such as shape and color of eyes, hair, mouth, etc.directly between persons while all other parts of the face remain unchanged. The network allows to generate modified images which appear like realistic images. Our model learns disentangled representations by weak supervision. We propose a localized resnet autoencoder optimized using several loss functions including a loss based on the semantic segmentation, which we interpret as masks, and a loss which enforces disentanglement by decomposition of the latent space into statistically independent subspaces. We evaluate the proposed solution w.r.t. disentanglement and generated image quality. Convincing results are demonstrated using the CelebA dataset.

AB - Image generating neural networks are mostly viewed as black boxes, where any change in the input can have a number of globally effective changes on the output. In this work, we propose a method for learning disentangled representations to allow for localized image manipulations. We use face images as our example of choice. Depending on the image region, identity and other facial attributes can be modified. The proposed network can transfer parts of a face such as shape and color of eyes, hair, mouth, etc.directly between persons while all other parts of the face remain unchanged. The network allows to generate modified images which appear like realistic images. Our model learns disentangled representations by weak supervision. We propose a localized resnet autoencoder optimized using several loss functions including a loss based on the semantic segmentation, which we interpret as masks, and a loss which enforces disentanglement by decomposition of the latent space into statistically independent subspaces. We evaluate the proposed solution w.r.t. disentanglement and generated image quality. Convincing results are demonstrated using the CelebA dataset.

KW - Autoencoders

KW - Face image editing

KW - Independent subspace analysis

KW - Latent space editing

KW - Machine learning

UR - http://www.scopus.com/inward/record.url?scp=85082453249&partnerID=8YFLogxK

U2 - 10.48550/arXiv.1908.08989

DO - 10.48550/arXiv.1908.08989

M3 - Conference contribution

AN - SCOPUS:85082453249

SN - 978-1-7281-5024-6

T3 - IEEE International Conference on Computer Vision Workshops

SP - 560

EP - 568

BT - 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 2019 IEEE/CVF 17th International Conference on Computer Vision Workshop (ICCVW)

Y2 - 27 October 2019 through 28 October 2019

ER -