BeautyGAN 阅读笔记

BeautyGAN 阅读笔记

@inproceedings{RN57,
author = {Li, Tingting and Qian, Ruihe and Dong, Chao and Liu, Si and Yan, Qiong and Zhu, Wenwu and Lin, Liang},
title = {Beautygan: Instance-level facial makeup transfer with deep generative adversarial network},
booktitle = {Proceedings of the 26th ACM international conference on Multimedia},
pages = {645-653},
type = {Conference Proceedings}
}

Contribution

这篇论文的主要contribution就是提出了 instance-level makeup transfer network BeautyGAN 和一个包含 makeup 和 non-makeup 图片的数据集。

Important Points

1. 采用衡量 face、lips、shadow 这三个地方的 pixel-level histogram loss；这里还要先对人脸进行分割，然后使用histogram matching对齐。
2. 提出了一个 3834 张图片的 makeup 数据集；

Motivation

之前提出的 CycleGAN 只能实现domain level的转换，makeup style 要求 instance level 的转换，而且很难获取到 pair of well-aligned face images with different makeup styles。