Object-driven Text-to-Image Synthesis via Adversarial Training (Obj-GAN; CVPR 2019)

[hcchiu1202]

0. 論文

タイトル: Object-driven Text-to-Image Synthesis via Adversarial Training
リンク: https://arxiv.org/abs/1902.10740
著者: Wenbo Li, Pengchuan Zhang, Lei Zhang, Qiuyuan Huang, Xiaodong He, Siwei Lyu, Jianfeng Gao
所属: University at Albany, SUNY, Microsoft Research AI, Microsoft, JD AI Research
発表年: 2019
掲載: CVPR

1. どんなもの？

First, generate a semantic layout (with class labels, bounding boxes, shapes of objects) from the text.
Then, using the both the layout and text as input, generate each object separately with attentions towards different levels, especially useful is the Object-driven attention.

2. 先行研究と比べてどこがすごい？

Taking the usual bi-LSTM encoded word/sentence as input, giving +27% Inception score and -11% FID score compared to AttnGAN.

3. 技術や手法のキモはどこ？

Generation of the lay-out without extra input/label.
The novel object-driven attentive generative network and the object-wise discriminator.
(Details to be updated shortly.)

4. どうやって有効だと検証した？

Will be updated shortly.

5. 議論はある？

Will be updated shortly.

6. 次に読むべき論文は？

Will be updated shortly.