1. 常用词的N种说法

1.1 Besides

Moreover
In addition

1.2 Get

obtain

1.3 Show

exhibit
demonstrate
present

1.4 Improve

be boosted to

1.5 Compare

In contrast

1.6 Can

allow to
attempt to
be able to
be exploited for
play the role of
be capable of

1.7 Many

diverse 多种多样的
multiple

1.8 Motivated by

Inspired by

1.9 Suppose

assume

1.10 Use

utilize
employ
apply … to …
enforce
be easily integrated with
deploy … as …
be taken as
be employed as

1.11 solve

address

1.12 define

refer to sth as sth
sth be identical to sth
be represented as

1.13 for example

for instance
take … as an example

1.14 come from

drift from

1.15 aim to

for the purpose of

1.16 without

dispensing with

2. 专有名词

2.1 Image-to-Image Translation

2.1.1 Word

stylised image 转换后的风格图
paired image 成对的图片
be transfered to
a fixed target style
multimodal translation 多迁移
deconvolution layer 反卷积层
stable training
uniform sampling
randomly sampled from

2.1.2 Advantage

deterministic one-to-one mapping transfer a soure image into the target style
these works can be divided into two categories according to the controllabiliy of the target styles
transfer the images in S into the style of T
transfer source images into the target style
By jointly optimizing all modules, CycleGAN model is able to transfer source images into the target sytle and v.v.
We take the S->T direction as an example, and the other direction can be similarly applied.
After the model is learnt, source images can only be translated to a fixed style

2.1.3 Disadvantage

Most previous studies for GAN-baed Image-to-Image translation methods rely on various forms of cycle-consistency$^{[1]}$.

2.2 Domain Adaption

2.2.1 Word

source domain 源域
target domain 目标域
intermediate domain 中间域
generalization ability 生成能力
domainness variable 邻域变量
pixel level
synthetic data 合成数据
real scenario 真实场景
be proportional to 和…成比例
recover … from …
be implemented with 被实施
spread along … from … to …
cross-domain semantic segmentation problem
a mixture of different target styles
shift from … to …
domain gap
domain alignment
distribution difference
semantic constraint 语义约束

2.2.2 Advantage

Domain adaptation aims to utilize a labeled source domain to learn a model that performs well on an unlabeled target domain
Domain generalization aims to learn a model that could be generalized to an unseen target domain by using multiple labeled source domains
align the image distributions for two domains
Object function of GAN can be seen as a lower bound of the Hessen-Shannon divergence
translate each source image into an arbitrary intermediate domain
Unsupersived domain adaption seeks to adapt the model trained on the source domain to the target domain$^{[1]}$.
Self-ensembling is composed of a teacher and a student network, where the student is compelled to produce consistent predictions provided by the teacher on target data$^{[1]}$.

2.2.3 Disadvantage

require a large amount of data with pixel-level annotations
The adversarial loss may trigger a negative transfer, which aligns the target feature with the source feature in an incorrect semantic category$^{[1]}$.

2.3 Style Transfer

2.3.1 Word

style patterns
the iterative optimization process
feedforward methods
insufficient visual quality
global feature statistics
in the feature space
encoder-decoder module
global color distribution
texture
local style patterns
bush strokes
receptive field

2.3.2 Sentence

Arbitrary style transfer aims to synthesize a content image with the style of an image to create a third image that has never been seen before.
The ultimate goal of arbitrary style transfer is to simultaneously achieve and preserve generation, quality, and efficiency.
We use the encoder(a pre-trained VGG-19) to compute the loss function for training …

3. 常用短语

draw increasing attention
most existing works
the benefit of
training phase 训练阶段
inference phase 测试阶段
be injected into 被注入
perform… by …
network structure
tent to
at the beginning
at the end
each of sth
computer vision tasks
achieve good performance
fill the gap between .. and …
conduct experiments on
training policy
verify the ability of
as expectedoop
give comparable result
by further using
from the first column, …
find it challenging to …
the seminal work of …的开创性工作
computational cost of …
computationally expensive operations
the proposed model
network structure
experimental settings
less appealing
run time performance
in real time
be compelled to 被迫

3.1 Solve the problem

To adress this challenging issue
To overcome this limitation

4. 常用衔接语

on one hand…, on the other hand
Unlike the above work
Our work is partialy inspired by
Formally, …
In particular, …
In this way, …
When …., …; and when, …
no longer aim to …, but to …
Specifically
Due to the usage of …
With regard to …
The motivation is as follows, …
Accordingly, …
In other words, …
In this section, …
It can be observed that
More interestingly, …
As discussed in
Note, …
In this paper, …
Experimental results demonstrate that …
Significant efforts have been made to …
Despite valuable efforts
Despite recent advances, …
In many cases, …
Similar to …

5. 长句型

Specifically, we introduce two discriminators, $D_S(x)$ to distinguish $M^{(z)}$ and S, and $D_T(x)$ to distinguish $M^{(z)}$ and T, respectively.
… in two settings. In the first setting, … . In the second setting, … .
Two cases are considered, … . For the first case, … . For the second case, … .
… in two scenarios. Firstly, … . Secondly, … .

References

Choi, Jaehoon & Kim, Taekyung & Kim, Changick. (2019). Self-Ensembling with GAN-based Data Augmentation for Domain Adaptation in Semantic Segmentation.