GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data
来源:互联网 发布:防蓝光夹片 知乎 编辑:程序博客网 时间:2024/06/15 15:48
本文的工作属于图像到图像间的翻译。类似于 DiscoGAN、CycleGAN 和 DualGAN,作者通过对偶学习,加上弱监督信息(weak 0/1 label),成功实现图像属性的迁移。G 采用 autoencoder 结构,encoder 将图像编码成图像主体信息(如:图像背景和人脸 ID 等)和属性信息(如:微笑、戴眼镜、发型等),decoder 则类似于 CGAN,将图像主体信息和属性信息翻译成图像。通过修改输入 decoder 的属性信息,实现属性的迁移。这种迁移成功的关键在于对偶学习机制(L1 重构误差),以及平行四边形 loss 和 nulling loss。值得一提的是,跟 CycleGAN 等相比,GeneGAN 只需要 generator 和 discriminator 各一个。文章在 CelebA 和 Multi PIE 数据集上进行实验,得到了不错的人脸属性迁移效果。 亮点推荐:推荐学习文章对弱监督信息的处理技巧。
代码地址:https://github.com/Prinsphield/GeneGAN
论文链接:https://arxiv.org/abs/1705.04932
推荐人:洪佳鹏,北京大学(PaperWeekly arXiv组志愿者)
阅读全文
0 0
- GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data
- 笔记:Learning Robust and Discriminative Subspace With Low-Rank Constraints
- Learning from imbalanced data
- AttributeError: 'Graph' object has no attribute 'cypher' in migration of data from Postgress to Neo4
- 前景提取--Subspace learning models
- Paper:Learning from Imbalanced Data
- Learning from data: Is Learning Feasible?
- AttributeError: 'Bunch' object has no attribute 'data'
- 【每周一文】Learning Classifiers from Only Positive and Unlabeled Data(2008)
- #Paper Reading# Topic Modeling using Topics from Many Domains, Lifelong Learning and Big Data
- .NET Custom Attribute Learning and Practice
- 读Learning belief networks from empirical data
- Basic Concepts for Learning from Data
- Self-taught Learning: TransferLearning from Unlabelled Data
- 《Learning From Data》读书笔记——引言
- 《Learning From Data》第一章(一)
- 《Learning From Data》第一章(二)读书笔记
- Learning from data: Training versus Testing
- 分享
- TP框架中的页面继承,以及控制器
- JVM中触发full gc的条件
- spark学习-39-Spark的StaticMemoryManager
- PHP学习笔记 15
- GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data
- C language actually comes with qsort function
- View控件可见性
- MFC学习笔记7 MFC基础
- 把后台的调试信息显示到浏览器控制台
- 表示数值的字符串
- Bootstrap 模态框垂直居中处理
- 如何判断Linux主机是否被攻击
- 万亿平安,如何继续高速增长?