Skip to content

Latest commit

 

History

History
27 lines (19 loc) · 1.3 KB

AL_CGAN.md

File metadata and controls

27 lines (19 loc) · 1.3 KB
Raw

#Learning to Generate Images of Outdoor Scenes from Attributes and Semantic Layouts

Main Idea

Propose a conditional generative adversarial network that generate outdoor scenes conditioning on semantic layout and scene attributes.

Pipeline

  • Generator: 5 convolutional layers + 4 deconvolutional layers
  • Descriminator: 6 convolutional layers + 1 fully connected layer
  • Both G and D see semantic condition and attributes condition

Key Points

  • The descriptive power of a generator can be increased by conditioning it with respect to side information, i.e. semantic map and attributes.
  • Different noise can yielding different generation, but from the images in this paper, the variant is tiny.

Dataset

TL;DR

This paper shows that providing additional side information in the form of conditioning variables is helpful for learning to generate better natural-looking images.

Reference

Levent Karacan, Zeynep Akata, Aykut Erdem, Erkut Erdem. Learning to Generate Images of Outdoor Scenes from Attributes and Semantic Layouts [J]. arXiv preprint arXiv:1612.00215, 2016.