0%

(ECCV 2018) Discriminative Region Proposal Adversarial Networks for High-Quality Image-to-Image Translation

Posted on In , ,

Wang C, Zheng H, Yu Z, et al. Discriminative region proposal adversarial networks for high-quality image-to-image translation[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 770-785.


1. Overview

论文提出Discriminative Region Proposal Adversarial Network (DRPAN)结构，并用于一些任务

  • de-raining
  • image-to-image translation

1.1. 模型



1.1.1. Discriminator

  • generate score map



  • find the most artificial patch by sliding window

  • map the patch into the real image, generate fake-mask real


  1. w. sliding window size
  2. w_s. score map size
  3. w_i. input image size
  4. w_*. patch size of input image that window map to
  5. x_c, y_c. center point of score map patch
  6. τ. scale
  7. d_r. discriminate region of input image

1.1.2. Reviser

  • discriminate

1.2. Loss Function



  • Discriminator



  • Reviser



    δ. noise

  • Generator



1.3. Algorithm



1.4. 数据集

  • Cityscapes
  • Facades
  • SUN6
  • UT Zappos 50K
  • edge to handbag
  • ID-CGAN
  • Pix2Pix
  • CRN (cascaded refined network)

1.6. Metric

  • PSNR
  • SSIM
  • VIF
  • RECO
  • IOU
  • FCN-score
  • AMT
  • Inception score

1.7. 实验结果