the ultimate goal of our approach is to use the mapped Source → Target samples ( ) to
augment the limited data of the target domain ( ).
let and be the task specific models trained on domains and .
domain is limited to only 10 samples per class, denoted
Tile is clear, "Few-Shot Adversarial Domain Adaptation"
DCD means domain-class discriminator(tells it's G1,G2,G3 or G4)
* FADA - n stands for our method when we use n labeled target samples per category in training*
The model trained on large data with more noise have more ability to adapte other domains.
A muti-domain image-to-image translation model.
Loss consist of Adversarial Loss(D tries to distinguish whether a image x from domain C is real or not, while G tries to cheat it), Domain Classification Loss (D tries to tell where a image is from, while tries to let D tell the fake image is from instead of ), Reconstruction Loss for Generator.
The structure is adapted from CycleGAN.
When trainning one domain, set other domains to be 0.
Doing the same Domain Adaptation job as CYCADA, translate images between GTA5 and Cityscapes.
This is published after CYCADA.
The main difference is that there's a Embedding space in the middle.
The loss consis of 6 parts:
In total, there's 3 discriminator and 2 encoder and 2 decoder and a classifier .
State of art performance when this paper was published.
In job above the encoder is LeNet.
Someting interesting: "Switching
To match the joint distribution with .
Find a matrix to let as close as possible.
So is .
For the first pair, we use a method called TCA to minimize
this equals to
For the next part, because we don't have , so we train a classfier C to learn
We have and , but are not from the same ditribution, i.e. . To solve this, we use the following method.
There exist a transformation that .
To find , we try to minimize following distance called MMD(maximum mean discrepancy).