==Browsing on Github is recommended==
[SRFlow: Learning the Super-Resolution Space with Normalizing Flow]
[2020] [ECCV]
Click to expand
**Research Problem:**
[Domain-Adversarial Training of Neural Networks(DANN)]
[2016] [JMLR] [📝] [
]
Click to expand
The main work:
To solve the problem of Domain Adaptation
[Multi-Task Adversarial Network for Disentangled Feature Learning]
[2018] [CVPR] [📝] []
Main idea:
Using a GAN like architecture to disentangles different features
[Disentangled Representation Learning GAN for Pose-Invariant Face Recognition]
[2017] [CVPR] []
Click to expand
key words:
Pose-Invariant Face Recognition (PIFR);
Problem:
solve the problem of pose-invariant face recognition (PIFR), the goal is to extract the identity representation that is exclusive or invariant to pose and other variations.
Related work:
existing PIFR methods can be group into two categories - 1) employ face frontalization to synthesize a frontal face and then use traditional face recognition methods. 2) learn features directly from the non-frontal face.
Impact:
help law enforcement practitioners identify suspects
Main method:
- the input of
$G_{enc}$ is a face image of any pose; the output is a feature representation.- the input of
$G_{dec}$ is a feature representation above, a pose code$c$ , and a random noise vector$z$ ; the output of$G_{dec}$ is a synthetic face image at a target pose$D$ is trained to not only distinguish real vs. fake images, but also predict the identity and pose of a face.
Mainly including three features:
- identity - represented by feature extracted by
$G_{enc}$ - pose - represented by pose code
- other facial feature (appearance variations) - represented by noise vector
The learning problem are twofold: 1) to learn a pose-invariant identity representation for PIFR, and 2) to synthesize a face image
The approach is to train a DR-GAN conditioned on the original image
Given a face image with label
[Transforming and Projecting Images into Class-conditional Generative Networks]
[2020] [ECCV] project page []
Click to expand
 **Key words:**
image-edit
Problem:
Most methods apply only to synthetic images that are generated by GANs in the first place.
只能修改GAN生成的图片，而不能直接修改现有图片。
In the real-world cases, people would like to edit their own image.
Related work:
Train a network for each separate image transformation (training time & model parameters)
Projection [Generative visual manipulation on the natural image manifold] [Neural photo editing with introspective adversarial networks] [Invertible conditional gans for image editing]
Main method:
by searching for an appropriate latent code, we project the image to the manifold of images produced by GANs.
HoloGAN: Unsupervised Learning of 3D Representations From Natural Images
Click to expand
Main method:
Traditional GANs learn to map a noise vector
$z$ directly to 2D features to generate images.HoloGAN learn to map a learnt 3D representation to the 2D image space.
[Defense-GAN: protecting classifiers against adversarial attacks using generative models]
[2018] [ICLR] [📝] []
Click to expand
The main work:
To solve the problem of classification which is vulnerable to adversarial perturbations: carefully crafted small perturbations can cause misclassification of legitimate images. I can archive it into the field of Machine deception. (small perturbations do not affect human recognition but machine classifier)
I can summarize their work as follows: given a picture with deception, GAN is used to generate the picture without deception, and finally classifier is used to classify.
They use the GD of reconstruction error ($ |G(\mathbf{z})-\mathbf{x}|_{2}^{2} $) to find optimal $ G(z) $
The methods it used:
- Several ways of attack: Fast Gradient Sign Method (FGSM), Randomized Fast Gradient Sign Method (RAND+FGSM), The Carlini-Wagner (CW) attack
- Lebesgue-measure
Its contribution:
They proposed a novel defense strategy utilizing GANs to enhance the robustness of classification models against black-box and white-box adversarial attacks
My Comments:
This work can be referred to using AE (Auto Encoder) for noise reduction. It’s just an easy application of GANs.