Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Toward a Controllable Disentanglement Network
(Submitted on 22 Jan 2020 (v1), last revised 20 Jun 2020 (this version, v3))
Abstract: This paper addresses two crucial problems of learning disentangled image representations, namely controlling the degree of disentanglement during image editing, and balancing the disentanglement strength and the reconstruction quality. To encourage disentanglement, we devise a distance covariance based decorrelation regularization. Further, for the reconstruction step, our model leverages a soft target representation combined with the latent image code. By exploring the real-valued space of the soft target representation, we are able to synthesize novel images with the designated properties. To improve the perceptual quality of images generated by autoencoder (AE)-based models, we extend the encoder-decoder architecture with the generative adversarial network (GAN) by collapsing the AE decoder and the GAN generator into one. We also design a classification based protocol to quantitatively evaluate the disentanglement strength of our model. Experimental results showcase the benefits of the proposed model.
Submission history
From: Zengjie Song [view email][v1] Wed, 22 Jan 2020 16:54:07 GMT (6017kb,D)
[v2] Thu, 4 Jun 2020 06:13:18 GMT (5992kb,D)
[v3] Sat, 20 Jun 2020 04:02:22 GMT (5992kb,D)
Link back to: arXiv, form interface, contact.