【美味蟹堡王09-30今日营业】论文学习笔记
为什么要写成博客呢？我觉得无论在做具体项目还是准备开题，还是要养成每天都多少翻翻论文的好习惯，保持活跃思维~所以就先粗略地以天为单位写成笔记的形式啦！自我监督~加油~
PairedCycleGAN: Asymmetric Style Transfer for Applying and Removing Makeup[paper]
Abstract
This paper introduces an automatic method for editing a portrait photo so that the subject appears to be wearing makeup in the style of another person in a reference photo. Our unsupervised learning approach relies on a new framework of cycle-consistent generative adversarial networks. Different from the image domain transfer problem, our style transfer problem involves two asymmetric functions: a forward function encodes example-based style transfer, whereas a backward function removes the style. We construct two coupled networks to implement these functions – one that transfers makeup style and a second that can remove makeup – such that the output of their successive application to an input photo will match the input. The learned style network can then quickly apply an arbitrary makeup style to an arbitrary photo. We demonstrate the effectiveness on a broad range of portraits and styles.
Introduction (Part)
This paper introduces a way to digitally add makeup to a photo of a person, where the style of the makeup is provided in an example photo of a different person (Figure 1). One challenge is that it is difficult to acquire a dataset of photo triplets from which to learn: the source photo, the reference makeup photo, and the ground truth output (which preserves identity of the source and style of the reference). Previous work on style transfer avoids the need for such a training set by defining the style and content loss functions based on deep features trained by neural networks [8, 16, 18]. While those approaches can produce good results for stylization of imagery in general, they do not work well for adding various makeup styles to faces. A second challenge, specific to our makeup problem, is that people are highly sensitive to visual artifacts in rendered faces. A potential solution is to restrict the stylization range so as to define a specific color transformation space (such as affine transformations), or so as to preserve edges [18, 19, 16]. Unfortunately, this approach limits the range of makeup, because many styles include features that would violate the edge preservation property such as elongated eyelashes or dark eye liner. Inspired by recent successful photorealistic style transfer based on generative adversarial networks (GANs), we take an unsupervised learning approach that builds on the CycleGAN architecture of Zhu et al. [26]. CycleGAN can transfer images between two domains by training on two sets of images, one from each domain. For our application, CycleGAN could in principle learn to apply a general make-you-look-good makeup to a no-makeup face, but it would not replicate a specific example makeup style. Thus, we introduce a set of problems where the forward and backward functions are asymmetric. Another example application would be transferring patterns from an example shirt to a white shirt, where the paired backward function could remove patterns from a shirt to make it white. Such forward (style transfer) functions require a source image and reference style as input, whereas the backward function (style removal) only takes the stylized image as input. Our approach relies on two asymmetric networks: one that transfers makeup style and another that removes makeup (each of which is jointly trained with an adversary). Application of both networks consecutively should preserve identity of the source photo (Figure 2). Finally, to encourage faithful reproduction of the reference makeup style, we train a style discriminator using positive examples that are created by warping the reference style face into the shape of the face in the source photo. This strategy addresses the aforementioned problem of a ground truth triplet dataset.
The principal contributions of this paper are:
(1) A feedforward makeup transformation network that can quickly transfer the style from an arbitrary reference makeup photo to an arbitrary source photo.
(2) An asymmetric makeup transfer framework wherein we train a makeup removal network jointly with the transfer network to preserve the identity, augmented by a style discriminator.
(3) A new dataset of high quality before- and after-makeup images gathered from YouTube videos.
Domain Generalization With Adversarial Feature Learning[paper]
Abstract
In this paper, we tackle the problem of domain generalization: how to learn a generalized feature representation for an "unseen" target domain by taking the advantage of multiple seen source-domain data. We present a novel framework based on adversarial autoencoders to learn a generalized latent feature representation across domains for domain generalization. To be specific, we extend adversarial autoencoders by imposing the Maximum Mean Discrepancy (MMD) measure to align the distributions among different domains, and matching the aligned distribution to an arbitrary prior distribution via adversarial feature learning. In this way, the learned feature representation is supposed to be universal to the seen source domains because of the MMD regularization, and is expected to generalize well on the target domain because of the introduction of the prior distribution. We proposed an algorithm to jointly train different components of our proposed framework. Extensive experiments on various vision tasks demonstrate that our proposed framework can learn better generalized features for the unseen target domain compared with state-of-the-art domain generalization methods.
introduction字有点多，题目看上去也有点大诶，因为正文本身字就不小，竟然没有小号字体哼╭(╯^╰)╮
今日新知识点~
Transfer learning；
Domain adaptation【传送门】
Domain generalization【传送门】