Info
- Title: Towards Instance-level Image-to-Image Translation
- Task: Image Translation
- Author: Zhiqiang Shen, Mingyang Huang, Jianping Shi, Xiangyang Xue, Thomas Huang
- Date: May 2019
- Arxiv: 1905.01744
- Published: CVPR 2019
Highlights & Drawbacks
- The instance-level objective loss can help learn a more accurate reconstruction and incorporate diverse attributes of objects
- A more reasonable mapping: the styles used for target domain of local/global areas are from corresponding spatial regions in source domain.
- A large-scale, multimodal, highly varied Image-to-Image translation dataset, containing ∼155k streetscape images across four domains.
Motivation & Design
Disentangle background and object style in translation process:
The framework overview:
Loss Design
The instance-level translation dataset and comparisons with previous ones:
Performance & Ablation Study
The authors compared results with baselines like UNIT, CycleGAN, MUNIT and DRIT, using LPIPS distance to measure the diversity of generated images.
A visualization for generated style distribution is also provided.
Code
Related
- Image to Image Translation(1): pix2pix, S+U, CycleGAN, UNIT, BicycleGAN, and StarGAN
-
Image to Image Translation(2): pix2pixHD, MUNIT, DRIT, vid2vid, SPADE and INIT
-
TransGaGa: Geometry-Aware Unsupervised Image-to-Image Translation - Wayne Wu - CVPR 2019
- InstaGAN: Instance-aware Image-to-Image Translation - Sangwoo Mo - ICLR 2019