Semantically Consistent Image-to-Image Translation for Unsupervised

Domain Adaptation

Stephan Brehm, Sebastian Scherer and Rainer Lienhart

Department of Computer Science, University of Augsburg, Universit

atsstr. 6a, Augsburg, Germany

Keywords:

Image Translation, Semi-supervised Learning, Unsupervised Learning, Domain Adaptation, Semantic

Segmentation, Synthetic Data, Semantic Consistency, Generative Adversarial Networks.

Abstract:

Unsupervised Domain Adaptation (UDA) aims to adapt models trained on a source domain to a new target

domain where no labelled data is available. In this work, we investigate the problem of UDA from a

synthetic computer-generated domain to a similar but real-world domain for learning semantic segmentation.

We propose a semantically consistent image-to-image translation method in combination with a consistency

regularisation method for UDA. We overcome previous limitations on transferring synthetic images to

real looking images. We leverage pseudo-labels in order to learn a generative image-to-image translation

model that receives additional feedback from semantic labels on both domains. Our method outperforms

state-of-the-art methods that combine image-to-image translation and semi-supervised learning on relevant

domain adaptation benchmarks, i.e., on GTA5 to Cityscapes and SYNTHIA to Cityscapes.

1 INTRODUCTION

The problem of domain adaptation from a synthetic

source domain to the real target domain is mainly

motivated by the cheap and almost endless possi-

bilities of automated creation of synthetic data. In

contrast, data from the real domain is often hard to

acquire. This is especially true for most types of

labelled data. A common problem in computer vision,

for which acquiring real labelled data is exceptionally

hard, is semantic segmentation. Semantic segmen-

tation requires annotations at the pixel level. These

annotations commonly need to be created manually in

a time-consuming process that also demands rigorous

and continuous focus from human annotators. Due

to this, a reasonable approach is to learn as much

as possible from synthetic data and then transfer the

knowledge to the real domain. However, convo-

lutional neural networks (CNNs), in general, learn

features from the domain on which they are trained

on. This causes networks to perform poorly on

unseen domains due to the visual gap between these

domains. Unsupervised domain adaptation (UDA)

aims to bridge this domain gap in order to learn

models that perform well in the target domain without

ever using labels from that domain.

In this work, we aim to improve the quality and

usefulness of synthetic data by transforming synthetic

images such that they look more like real images.

However, such an Image-to-Image Translation (I2I)

approach cannot change fundamental differences in

image content because it needs to keep the trans-

formed images consistent to the semantic labels of

synthetic images. Because of this, we cannot bridge

the gap between synthetic and real domains solely by

an I2I approach.

The remaining differences are in image content

and include object shapes, object frequency, and

differences in viewpoint. Following recent work, we

utilise a Semi-Supervised Learning (SSL) framework

which allows us to include unlabelled data from the

real target domain into the training of a semantic

segmentation network. The main contributions of our

work are summarised as follows:

1. We propose a semantically consistent I2I method

that combines an adversarial approach with a

segmentation objective that is optimised jointly by

both generator and discriminator.

2. We improve both adversarial and segmentation

objectives with self-supervised techniques that

include the unlabelled data from the real target

domain into the training. Our method outperforms

state-of-the-art combinations of I2I and SSL on

challenging benchmarks for UDA. We provide ex-

tensive experiments and analyses and show which

components are essential to our approach.

Brehm, S., Scherer, S. and Lienhart, R.

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation.

DOI: 10.5220/0010786000003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 2, pages 131-141

ISBN: 978-989-758-547-0; ISSN: 2184-433X

131

2 RELATED WORK

Semantic Segmentation: is the task of labelling

every pixel of an image according to the object class

it belongs to. In 2015, Long et al. proposed fully

convolution neural networks (FCNs) (Long et al.,

2015) improving massively upon classical methods

for semantic segmentation. Since then, many new

methods based on deep convolutional neural networks

were proposed (Chen et al., 2018; Wang et al., 2021;

Yu and Koltun, 2016). However, these methods

are directly learned on the real target domain in a

supervised fashion. In contrast, our method is able

to learn about the real domain without the necessity

of annotated real data.

Image-to-Image Translation (I2I): methods (Choi

et al., 2019; Pizzati et al., 2020) have been widely

used to bridge the gap between synthetic and real

data. For image data, this is commonly achieved

by learning a deep convolutional neural network that

receives a synthetic image from the source domain

as input and manipulates it in such a way that it

looks more realistic. Basically, these methods try

to make the synthetic data look more real. For our

task, it is important that the manipulated versions of

the synthetic data can be used for supervised training

of a segmentation method subsequently. In order

to learn useful manipulations for such a task, these

systems need mechanisms that keep the overall image

consistent with the synthetic labels.

Unsupervised Domain Adaptation (UDA): for se-

mantic segmentation has been extensively studied in

the last years. Adversarial training is often used in

UDA methods to adapt either input space, feature

space or output space of a semantic segmentation

network (Toldo et al., 2020). On the feature level, a

discriminator is trained to distinguish between feature

maps from different domains (Hoffman et al., 2016;

Chen et al., 2017; Hong et al., 2018). Popular

input space adaptation techniques utilise frameworks

based on cycle consistency (Zhu et al., 2017) for

UDA (Hoffman et al., 2018; Sankaranarayanan et al.,

2018; Chen et al., 2019; Murez et al., 2018). How-

ever, cycle consistency allows almost arbitrary trans-

formations as long as they can be reversed. In contrast

to these methods, we perform input space adaptation

in our I2I method without the need for cycle consis-

tency. Our method simply uses the annotations of the

synthetic data to directly enforce consistency.

Semi-Supervised Learning (SSL): aims to include

unlabelled data into the training which allows to use

data from the real domain. The dominant approaches

for SSL are pseudo-labelling and consistency regu-

larisation. An extensive overview can be found in

this survey (van Engelen and Hoos, 2020). Pseudo-

Labeling was ﬁrst proposed in (Lee, 2013). Xie et

al. (Xie et al., 2020) recently showed that pseudo-

labels can indeed improve overall performance on

image classiﬁcation tasks. They train a network on

labelled data and reuse high-conﬁdence predictions

on unlabelled data as pseudo-labels. Pseudo-labels

are then included in the full dataset on which a new

model is subsequently trained from scratch.

Many recent SSL methods (Ke et al., 2019; Tar-

vainen and Valpola, 2017) include consistency regu-

larisation. They employ unlabelled data to produce

consistent predictions under different perturbations.

Possible perturbations can be data augmentation,

dropout or simple noise on the input data. The trained

model should be robust against such perturbations.

These approaches leverage the idea that a classi-

ﬁer should output the same distribution for different

augmented versions of an unlabelled sample. This

is typically achieved by minimising the difference

between the prediction of a model on different per-

turbed versions of the same input. We utilise such an

approach to boost our performance even further. Our

I2I translation module also leverages pseudo-labels

created by such a consistency regularised method.

3 METHOD

In this work, we assume that image data is available

from both the synthetic source domain as well as

the real target domain. However, annotations are

only available in the synthetic source domain. We

aim to bridge the gap between the source and the

target data. Our image-sets are sampled from their

respective domains. The ﬁrst step is to bring the

two domains closer together visually which means

transforming the synthetic data such that it looks

more real. By keeping consistency with the labels

of the synthetic source data, we are able to generate

a dataset of images with corresponding labels that

closely resembles the real target data in terms of

colour, texture and lighting. Basically, we aim to

improve domain adaptation by transferring the style

of the real target domain onto images of the synthetic

source domain. However, fundamental differences in

image composition cannot be learned with such an

I2I method. For this reason, we use the translated

synthetic data in conjunction with unlabelled data

from the real target domain to train our segmentation

model. We propose a training process that consists of

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

132

three phases.

• (a) In the warm-up phase the SSL method is used

to train an initial segmentation model M

as the

initial pseudo label generator.

• (b) In the I2I training phase, we use the pseudo

labels from model M

and train our generator G

to produce real looking images from synthetic

images.

• (c) In the segmentation training phase, we com-

bine the SSL method with the translated images

by generator G and train our new segmentation

network M

3.1 Notation

In the following, we will denote S as the source

domain and T as the target domain. X

and X

are

sets of images sampled from S and T , i.e., synthetic

images and real images, respectively. x

∈ X

and

∈ X

are images sampled from their respective

image sets X

and X

Both domains S and T share a common set of

categories C. We denote the label of an image x

domain d ∈ {S , T } as y

. In this work y

∈ Y

gener-

ally is a synthetic segmentation mask and

∈

a segmentation mask that contains pseudo-labels. For

the purpose of this work, the real segmentation labels

∈ Y

are not available for training. L

W,n

i, j,c

(x, y)

denotes a loss function based on image x and label

y that is used to update trainable parameters of a

network W . n is an identiﬁer/name for the loss

function. Note that we often omit the arguments

(x, y) in order to simplify our notation. Also note

that W only indicates which network is updated by

and thus may also be omitted in more general

statements and deﬁnitions. i, j, c are indices that we

use to refer to speciﬁc dimensions and or individual

values in the calculated loss if necessary. In general,

we assume multidimensional arrays to be in the order

of height × width × channels.

In this work, we aim to train a segmentation

network F

on X

(and possibly X

) that estimates

without the need for any y

∈ Y

during training,

i.e., we aim to generalise from domain S to domain T

without ever using any labels from T .

3.2 Image-to-Image Translation

In order to generalise to a target domain T we need

to bring our source images X

closer to T . In this

section, we detail our approach to this goal. Figure 1

illustrates the employed architecture. We aim to

transfer images from a synthetic domain to the real

domain, because the synthetic domain allows to easily

create images and corresponding labels by simulating

an environment that is similar to the real environment.

The transformed images are supposed to deliver max-

imum performance on real data when learning a fully

convolutional method for semantic segmentation. In

order to achieve this, such a transformation needs to

meet two major requirements.

1. Transformed images need to look as realistic as

possible

2. Transformed images need to maintain consistency

to the segmentation labels

We propose an I2I method that consists of two net-

works in an adversarial setting. Our learning task

is designed to enforce the requirements given above.

Requirement 1. is tackled by an adversarial objec-

tive. For Requirement 2., we extend the adversarial

setting with an additional cooperative segmentation

objective that is jointly optimised by both adversaries.

Following, we give a brief overview of the employed

generator and discriminator architectures.

Generator. We use a simple encoder-decoder archi-

tecture with strided convolutions for down-sampling

the input image by a factor of 8. These down-

sampling operations are followed by a single residual

block. The decoder is simply a stack of deconvolution

layers. Note that we do not include any skip/residual

connections between encoder and decoder. In every

layer, we append a learned scaling factor as well

as a learned bias term. We use leaky-ReLU acti-

vations for all layers except the output layer. We

also use deconvolution kernels, PixelNorm, Equalized

Learning Rate, Adaptive Instance Normalization and

Stochastic Variation as proposed in (Karras et al.,

2018). Similar to Taigman et al. (Taigman et al.,

2017), we incorporate images from the target domain

in the training of the generator by adding an identity

objective. Thus, the generator learns to encode and

reconstruct real images in addition to its main I2I task.

The additional feedback is a way to directly learn

about the structure and texture of real images.

Discriminator. Instead of discriminating between

generated images and real images directly, we pro-

pose to use prior knowledge for discrimination. We

use image-features of an ImageNet-pretrained (Deng

et al., 2009) VGG16 network (Simonyan and Zisser-

man, 2015) after the block3 conv3 layer. We build

additional 3 residual blocks (He et al., 2016) on top

of these features. Note that we do not ﬁne-tune the

pre-trained VGG16 part of the discriminator. Our

discriminator features two distinct outputs which are

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation

133

synthetic label

pseudo label

synthetic image

real image

Identity

Loss

Generator

Segmentation

Loss

GAN-Loss +

Classwise

GAN-Loss

Discriminator

Figure 1: Schematic illustration of our proposed I2I method. The colored arrows visualize data-ﬂow.

each computed on top of a separate decoder, as illus-

trated in Figure 1. We refer to these decoders as the

GAN-head (Goodfellow et al., 2014) and the auxiliary

classiﬁer head (AC-head) (Odena et al., 2017). Both

heads output results at image resolution. The task of

the AC-head is to learn semantic segmentation in the

target domain. Thus we feed the translated images

in conjunction with the labels of the synthetic data

as well as real images with corresponding pseudo-

labels. Note that these pseudo-labels can be created

online during training. However, using pseudo-labels

created by our proposed segmentation system that

is detailed in section 3.3 generally improved results.

The AC-head is important in many ways. First,

the segmentation feedback generated in the AC-head

forces previous layers to learn segmentation-speciﬁc

features which in turn helps the GAN-head to dis-

criminate between images that are translated from the

synthetic domain and images from the real domain.

Second, it provides consistency feedback for the

generator. We achieve this by propagating gradients

based on the segmentation loss of the AC-head into

the weights of the discriminator as well as into the

weights of the generator. This effectively optimises

a joint segmentation objective in both generator and

discriminator. Thus, the generator is punished if the

translated images are not consistent with the label

of the synthetic input data. Third, it can be used to

generate pseudo-labels. The GAN-Head is responsi-

ble for the discrimination between real examples and

fake examples. Note that GAN-Heads are commonly

learned with a single label only (real or fake). We

extend this formulation to include class information.

Thus, allowing the discriminator to directly learn to

compare images on a class speciﬁc level. We utilize

segmentation labels in the GAN-head to create this

class-speciﬁc feedback. In GAN-terminology this

means that we treat pseudo-labels as real and the

synthetic labels as fake. For each class, we produce

an additional output feature map. On this feature map,

feedback is only applied at positions that belong to the

given object class as deﬁned by either synthetic labels

or pseudo-labels.

Optimisation. Given the supervised segmenta-

tion loss L

seg

(D(G(x

)), y

) (softmax cross-entropy)

based on synthetic labels and the supervised segmen-

tation loss L

seg

(D(x

) based on the pseudo-labels,

we compute the total segmentation loss L

seg

as given

in Equation 1. We use the symmetric cross-entropy

loss (Wang et al., 2019) for L

seg

(D(x

) because it

increases robustness when using noisy pseudo-labels

seg

= L

seg

(D(G(x

)), y

) + λ

seg

(D(x

) (1)

where λ

is a scaling factor that we use to control

the relative impact of the pseudo-labels on the overall

error of the model.

Like the segmentation loss, we calculate the

GAN-Loss L

gan

on a pixel-wise basis, i.e., the GAN-

loss is the average of all losses computed over all pix-

els I

i, j

of an input image I. Let m(a, b) be a derivable

distance metric between a and b. We give a general

error function L

D,dgan

for a discriminator D as well as

a general error function L

G,dgan

for a generator G in

Equation 2 and Equation 3 respectively.

D,dgan

i, j

= m(D(G(x

))

i, j

, 0) + m(D(x

)

i, j

, 1) (2)

G,dgan

i, j

= m(D(G(x

))

i, j

, 1) (3)

We use the mean-squared error for the distance metric

Now let, y

be the one-hot encoded segmentation

label for a given image x

, i.e., y

i, j,c

equals one if I

i, j

belongs to class c. Otherwise y

i, j,c

is zero. The same

applies to the pseudo-label

. We deﬁne a class-wise

and pixel-wise error L

D,cgan

i, j,c

for discriminator D in

Equation 4 and a class-wise and pixel-wise error

G,cgan

i, j,c

for generator G in Equation 5.

D,cgan

i, j,c

= m(D(G(x

))

i, j,c

, 0) · y

i, j,c

+ (4)

+ m(D(x

)

i, j,c

, 1) ·

i, j,c

G,cgan

i, j,c

= m(D(G(x

))

i, j,c

, 1) · y

i, j,c

(5)

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

134

Hence, the full adversarial loss for D is:

D,gan

∑

i=0

∑

j=0

D,dgan

i, j

cgan

|C|

∑

c∈C

D,cgan

i, j,c

. (6)

where λ

cgan

is a scaling factor that we use to control

the relative impact of the class-wise loss on the overall

error of the model, H is the height and W is the width

of both images and labels. L

G,gan

can be computed

analogous.

The total loss of the discriminator L

is a com-

bination of segmentation and GAN losses as given in

Equation 7.



D,seg

+ L

D,gan



(7)

With L

G,id

being the identity reconstruction error on

a real image x

from the target domain, we compute

the generator loss L

as given in Equation 9.

G,id

∑

i=0

∑

j=0

kG(x

)

i, j

− x

i, j

(8)



G,seg

+ L

G,gan

+ L

G,id



(9)

Note that the second term of L

seg

from Equation 1

is not affected by G and thus L

G,seg

is reduced to

the ﬁrst term of L

seg

when back-propagating. This

means that G should generate images that match the

synthetic segmentation labels. On the other hand,

G,gan

incentivises more realistic images. By using

G,id

we are able to show real images from the target

domain to G. This combination of various objectives

enables us to learn an I2I model that satisﬁes the

requirement of realistic transformation while simulta-

neously keeping consistency with the synthetic labels.

We use the resulting images to subsequently learn a

better segmentation model.

3.3 Segmentation Training

At this point, we are able to transform synthetic

images into real images while keeping the semantic

content of the image unchanged. Thus, we can use the

transformed images with the original synthetic anno-

tations. However, some issues remain to be solved.

An I2I method cannot bridge certain differences be-

tween the domains. The remaining differences are

in image content and include object shapes, object

frequency, and differences in viewpoint which can

not be learned by our I2I module. As a solution,

we incorporate images from the real domain directly

into the training of the segmentation model via a SSL

framework.

Similar to Tarvainen et al. (Tarvainen and Valpola,

2017), we make use of two networks: a student net-

work F

and a teacher network F

. The architecture

of the teacher network is identical to the one of the

student network. The weights of the teacher model

are an Exponential Moving Average (EMA) of the

student’s weights. Given an image from the target

domain, the teacher’s prediction serves as a label for

the student, forcing consistency in the prediction of

both models under different perturbations.

The overall objective L

is a combination of

a supervised segmentation loss L

,seg

and the self-

supervised consistency loss L

,con

as detailed in

Equation 10.

= L

,seg

, y

) + λ

con

,con

), (10)

where λ

con

is a trade-off parameter.

For the supervised training, we incorporate syn-

thetic images x

∈ X

and a transformed version

obtained from the generator G(x

) of our proposed

I2I method. We calculate a combined loss L

,seg

given in Equation 11

,seg

= L

seg

((F

)), y

) + L

seg

((F

(G(x

))), y

(11)

where L

seg

again is the softmax cross-entropy error.

For the semi-supervised training, we incorporate

images from the real domain x

∈ X

. We calculate

the consistency loss L

,con

) as given in Equa-

tion 12.

,con

= ||σ(F

)) − σ(F

(P(x

)))||

, (12)

where σ is the softmax activation function and P(x

)

is a strongly perturbed version of the input image

. Similar to (Zhou et al., 2020), we utilise color

jittering, Gaussian blurring and noise as perturbations

on x

As our I2I model can also leverage pseudo-labels,

we perform an identical training with a supervised

loss and a consistency loss, but without transformed

images G(x

) in the warm-up phase. We use the

resulting model to generate pseudo-labels that we use

to train G in the I2I training phase. We then use G

to translate synthetic images. These translated images

in conjunction with the original synthetic images con-

stitute the training data for the segmentation training

phase.

4 EXPERIMENTS AND RESULTS

In this section, we detail experiments that we con-

ducted in order to show the performance of our

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation

135

Table 1: Results of domain adaptation from GTA5 to Cityscapes using a VGG backbone.

road

sidewalk

building

wall

fence

pole

trafﬁc light

trafﬁc sign

vegetation

terrain

sky

person

rider

car

truck

bus

train

motorbike

bicycle

mIoU

Source only (ours) 79.2 30.6 76.0 22.7 11.1 19.2 11.0 2.2 80.3 30.4 73.5 39.3 0.5 75.3 17.3 9.5 0.0 1.4 0.0 30.5

CyCADA (Hoffman et al., 2018) 85.2 37.2 76.5 21.8 15.0 23.8 22.9 21.5 80.5 31.3 60.7 50.5 9.0 76.9 17.1 28.2 4.5 9.8 0.0 35.4

CBST-SP (Zou et al., 2018) 90.4 50.8 72.0 18.3 9.5 27.2 28.6 14.1 82.4 25.1 70.8 42.6 14.5 76.9 5.9 12.5 1.2 14.0 28.6 36.1

DCAN (Wu et al., 2018) 82.3 26.7 77.4 23.7 20.5 20.4 30.3 15.9 80.9 25.4 69.5 52.6 11.1 79.6 24.9 21.2 1.3 17.0 6.7 36.2

SWD (Lee et al., 2019) 91.0 35.7 78.0 21.6 21.7 31.8 30.2 25.2 80.2 23.9 74.1 53.1 15.8 79.3 22.1 26.5 1.5 17.2 30.4 39.9

SIM (Wang et al., 2020b) 88.1 35.8 83.1 25.8 23.9 29.2 28.8 28.6 83.0 36.7 82.3 53.7 22.8 82.3 26.4 38.6 0.0 19.6 17.1 42.4

TGCF-DA + SE (Choi et al., 2019) 90.2 51.5 81.1 15.0 10.7 37.5 35.2 28.9 84.1 32.7 75.9 62.7 19.9 82.6 22.9 28.3 0.0 23.0 25.4 42.5

FADA (Wang et al., 2020a) 92.3 51.1 83.7 33.1 29.1 28.5 28.0 21.0 82.6 32.6 85.3 55.2 28.8 83.5 24.4 37.4 0.0 21.1 15.2 43.8

PCEDA (Yang et al., 2020) 90.2 44.7 82.0 28.4 28.4 24.4 33.7 35.6 83.7 40.5 75.1 54.4 28.2 80.3 23.8 39.4 0.0 22.8 30.8 44.6

Zhou et al. (Zhou et al., 2020) 95.1 66.5 84.7 35.1 19.8 31.2 35.0 32.1 86.2 43.4 82.5 61.0 25.1 87.1 35.3 46.1 0.0 24.6 17.5 47.8

Ours 94.4 65.3 85.9 39.0 22.2 35.4 39.1 37.3 86.7 42.3 88.1 62.7 36.2 87.6 33.8 45.0 0.0 26.5 24.2 50.1

Target only (ours) 96.9 79.7 89.6 46.0 47.6 47.4 55.9 64.5 90.0 60.7 91.7 72.4 49.1 92.4 56.7 75.4 54.0 51.0 69.4 68.0

Table 2: Results of domain adaptation from SYNTHIA to Cityscapes using a VGG backbone. mIoU

and mIoU

are

computed on 16 and 13 classes respectively. * means that classes are included in mIoU

but excluded in mIoU

road

sidewalk

building

wall*

fence*

pole*

trafﬁc light

trafﬁc sign

vegetation

sky

person

rider

car

bus

motorbike

bicycle

mIoU

Source only (ours) 42.0 19.6 60.4 6.3 0.1 28.3 2.1 10.3 76.2 76.0 44.3 7.2 62.5 14.8 3.2 10.6 29.0 33.0

DCAN (Wu et al., 2018) 79.9 30.4 70.8 1.6 0.6 22.3 6.7 23.0 76.9 73.9 41.9 16.7 61.7 11.5 10.3 38.6 35.4 -

CBST (Zou et al., 2018) 69.6 28.7 69.5 12.1 0.1 25.4 11.9 13.6 82.0 81.9 49.1 14.5 66.0 6.6 3.7 32.4 35.4 40.7

SWD (Lee et al., 2019) 83.3 35.4 82.1 - - - 12.2 12.6 83.8 76.5 47.4 12.0 71.5 17.9 1.6 29.7 - 43.5

FADA (Wang et al., 2020a) 80.4 35.9 80.9 2.5 0.3 30.4 7.9 22.3 81.8 83.6 48.9 16.8 77.7 31.1 13.5 17.9 39.5 46.0

TGCF-DA + SE (Choi et al., 2019) 90.1 48.6 80.7 2.2 0.2 27.2 3.2 14.3 82.1 78.4 54.4 16.4 82.5 12.3 1.7 21.8 38.5 46.6

PCEDA (Yang et al., 2020) 79.7 35.2 78.7 1.4 0.6 23.1 10.0 28.9 79.6 81.2 51.2 25.1 72.2 24.1 16.7 50.4 41.1 48.7

Zhou et al. (Zhou et al., 2020) 93.1 53.2 81.1 2.6 0.6 29.1 7.8 15.7 81.7 81.6 53.6 20.1 82.7 22.9 7.7 31.3 41.5 48.6

Ours 94.8 67.2 81.9 6.1 0.1 29.6 0.1 19.7 82.2 81.1 50.2 17.0 84.6 30.8 12.4 25.1 42.7 49.8

Target only (ours) 96.9 79.7 89.6 46.0 47.6 47.4 55.9 64.5 90.0 91.7 72.4 49.1 92.4 75.4 51.0 69.4 70.0 75.4

proposed methods. In Section 4.1 we shortly intro-

duce the datasets that we used. In Section 4.2 we

give important details on the training and validation

protocols. In Section 4.3 we compare our methods

to previous state-of-the art methods on two public

benchmarks. In Section 4.4 we conduct an ablation

study to show the impact of individual components

on our results.

4.1 Datasets

Following common practice for unsupervised domain

adaptation in semantic segmentation, we use the

GTA5 (Richter et al., 2016) and SYNTHIA (Ros

et al., 2016) datasets as our synthetic domain and the

Cityscapes dataset (Cordts et al., 2016) as our real

domain. The datasets are detailed below.

Cityscapes: dataset (Cordts et al., 2016) contains

images of urban street scenes collected around Ger-

many and neighbouring countries. It consists of a

training set with 2975 images and a validation set

of 500 images. We report mIoU

for the 19 object

classes that are annotated. We also use the available

89250 unlabeled images for learning our I2I method.

GTA5: dataset (Richter et al., 2016) contains im-

ages rendered by the game Grand Theft Auto 5. It

consists of 24966 images with corresponding pixel-

level semantic segmentation annotations and a set of

object classes that is compatible to the annotations of

the Cityscapes dataset.

SYNTHIA: dataset (Ros et al., 2016) consists of

a collection of images rendered from a virtual city.

We use the SYNTHIA-RAND-CITYSCAPES subset,

which consists of 9400 images with pixel-wise anno-

tations. Note that the terrain, truck, train classes are

not annotated in the SYNTHIA dataset. Thus, we use

the remaining 16 classes that are common with the

Cityscapes dataset. We evaluate mIoU

on these 16

classes and mIoU

on a subset of 13 classes. mIoU

is a common metric for evaluation on the SYNTHIA

dataset that excludes certain classes that are especially

hard and/or underrepresented in the data.

4.2 Implementation Details

For a fair comparison to previous work (Choi et al.,

2019; Zhou et al., 2020), we adopt the VGG16 back-

bone (Simonyan and Zisserman, 2015) pre-trained

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

136

on the ImageNet dataset (Deng et al., 2009). Fol-

lowing Deeplab-v2 (Chen et al., 2018), we incor-

porate Atrous Spatial Pyramid Pooling (ASPP) as

the decoder and use an bi-linear up-sampling to get

the segmentation output at image resolution. We

use this model for all our experiments. We use

the Adam optimiser (Kingma and Ba, 2015) for the

segmentation training with an initial learning rate of

1 × 10

−5

and exponential weight decay. The gen-

erator and discriminator are trained with a constant

learning rate of 1 × 10

−5

for 1 million iterations.

For validation purposes, we keep exponential moving

averages of the generator weights. We use a 50/50

split of synthetic and transformed data to learn the

segmentation model. All networks are trained with

gradient clipping at a global norm of 5. We set all

EMA decay values to 0.999. During the ﬁrst 10.000

training steps of F

, we keep λ

con

to zero. We report

results of the teacher network which is a smoothed

version of the trained segmentation network. We

fade λ

pseudo

and λ

cgan

linearly from zero at iteration

20,000 to 0.3 at iteration 100,000 when learning

with pseudo-labels created online. When learning

with pre-computed pseudo-labels from our proposed

segmentation model, we keep λ

pseudo

and λ

cgan

0.3 at all times. All images in all experiments are

re-scaled such that the longer side is 1024 pixels.

For training, we crop 384 × 384 regions from these

images.

4.3 Comparisons with the

State-of-the-Art

We compare the results of our method to state-of-the-

art methods that combine I2I and SSL on the two

standard benchmarks: “GTA5 → Cityscapes” and

“SYNTHIA → Cityscapes” in Table 1 and Table 2,

respectively. Our method improves upon current

state-of-the art in both benchmarks. Like many other

methods, we suffer from fundamental differences in

the deﬁnition of certain objects, e.g., the train class

in the Cityscapes dataset which includes objects like

street-cars is actually very similar to the bus class

in the GTA5 dataset (Zhang et al., 2017). Such a

difference obviously impacts performance on both

classes. Indeed, due to the similarity of buses in

the source data to trains in the target data, we would

actually expect the converted buses to incorporate

certain features of the target trains. During segmenta-

tion training, we then actively enforce predictions for

the bus class which ultimately results in poor perfor-

mance. We identiﬁed an additional issue that affects

the car and truck annotations in the GTA5 dataset.

Here, one of the synthetic car models is consistently

Table 3: Ablation study for the proposed image-to-image

translation method from the synthetic GTA5 dataset to

the Cityscapes dataset. We report mean mIoU

on the

Cityscapes validation set.

Component mIoU

LSGAN (Mao et al., 2017) 43.9

SGAN (Goodfellow et al., 2014) 42.9

w/o L

D,G,clsgan

42.7

w/o L

G,seg

38.1

w/o pre-computed pseudo-labels 40.0

annotated with the truck label. We think that our

method that directly enforces label consistency is

overly prone to such differences and errors in the

annotations. Also note that domain adaptation from

SYNTHIA to Cityscapes, in general, is inferior to the

variant trained on the GTA5 dataset. The domain gap

between the SYNTHIA dataset and the Cityscapes

dataset is simply wider than the gap between GTA5

and Cityscapes. This is mainly due to the fact that

the viewpoint in the scenes in the SYNTHIA dataset

varies a lot while Cityscapes and GTA5 mainly show

scenes from the viewpoint of a pedestrian or the

viewpoint of a driver. The SYNTHIA dataset also

consists of less annotated images. Common practice

excludes many of the hard classes which results in

mIoU

scores being very similar to the mIoU

scores that we obtain when training with the GTA5

dataset. The above-mentioned issues apply to most

methods. However, our method outperforms other

methods. Thus, in the next section, we conduct

an ablation study to identify the contribution of the

individual components of our method to this outper-

formance.

4.4 Ablation Study

In this section, we conduct an ablation study on

various components of our proposed method. In

addition, we compare different components directly

to their counterparts in similar methods (Choi et al.,

2019; Zhou et al., 2020). We also analyse the

remaining domain gap to get a better understanding

of our results.

Image-to-Image Translation. The results of the

ablation study are shown in Table 3. For simplicity,

we omit the SSL part of the third stage of our method,

i.e., we use the images from our I2I method to learn

a segmentation model in a purely supervised fashion.

We train on 50% source data and 50% translated data.

As expected, our performance is highly dependent on

the semantic consistency loss L

G,seg

from Equation 9

that we use to supervise the generator. We lose ≈

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation

137

Table 4: Ablation study and comparison to existing similar

work. We evaluate and compare the image translation and

constancy regularization. We report results from GTA5

to Cityscapes. L

,seg

is the segmentation loss deﬁned in

Equation 11. SSL refers to the consistency loss L

,con

deﬁned in Equation 12. I2I means that we use translated

images from our proposed I2I method.

Method Component mIoU

Source only (ours) L

,seg

30.5

(Choi et al., 2019) L

,seg

+ SSL 32.6

(Zhou et al., 2020) L

,seg

+ SSL 35.6

Ours L

,seg

+ SSL 39.2

(Choi et al., 2019) L

,seg

+ I2I 35.4

(Zhou et al., 2020) L

,seg

+ I2I 35.1

Ours L

,seg

+ I2I 43.9

(Choi et al., 2019) L

,seg

+ SSL + I2I 42.5

(Zhou et al., 2020) L

,seg

+ SSL + I2I 47.8

Ours L

,seg

+ SSL + I2I 50.1

5.7% mIoU

absolute performance by removing this

component which demonstrates the effectiveness of

our approach. Note that this ablation evaluates the im-

pact of L

G,seg

. L

D,seg

is still applied during training,

i.e., the discriminator has access to the segmentation

information but is not able to properly transfer this

information to the generator. This can be improved by

applying L

G,seg

during training which explicitly en-

forces the transfer of segmentation knowledge. Note

that our method is still able to deliver 40% mIoU

when using simple pseudo-labels created online dur-

ing training from the output of the AC-head, i.e., if we

train without pseudo-labels supplied by an external

method. Nevertheless, using higher quality pseudo-

labels from the proposed segmentation method in-

creases performance by an absolute 3.8% mIoU

Removing the class-wise GAN feedback reduces per-

formance by around 1.2% mIoU

. Swapping the

LSGAN target to a standard GAN (SGAN) target

reduces performance by ≈ 1% mIoU

Figure 2 shows examples of transformed synthetic

images. We observe that the textures of roads and

trees look much more realistic. Also, the sky is

generally more cloudy which is very characteristic for

the Cityscapes dataset.

Comparison with Similar Methods. In Table 4,

we compare our two main components to the main

components of similar work by Choi et al. (Choi et al.,

2019) and Zhou et al. (Zhou et al., 2020). We com-

pare the components in isolation and in combination.

Both components improve performance substantially

upon previous work. We can clearly see that the

improvement of the I2I method is predominantly

achieved through the semantic consistency frame-

Table 5: Domain gap evaluation. Our method closes the

domain gap between GTA5 and Cityscapes by 68.3%.

mIoU

domain gap

Cityscapes Model 68.0 0.0%

Source only 39.6 100.0%

Ours 59.0 31.7%

work that we described in subsection 3.2. However, in

Table 3, we can also clearly see, that the performance

of our I2I method is heavily impacted by the quality of

the pseudo-labels. This shows that both components

beneﬁt each other.

4.5 Domain Gap Analysis

In order to estimate the remaining domain gap we

retrain the linear classiﬁcation layer of the segmenta-

tion model on real labels from the Cityscapes dataset.

We compare to a model that is trained on Cityscapes

only. The results are summarised in Table 5. We argue

that retraining the classiﬁcation layer is necessary for

a proper comparison because it allows to overcome

fundamental differences in class annotations between

the synthetic source and the real target data. In

essence, this means that we evaluate the quality of the

learned features and their applicability to data from

the target domain. In this context, feature quality

refers to the linear separability of the object classes

from the Cityscapes dataset in the features. This

linear separability can be assessed by training a linear

classiﬁer on top of these features. Such an approach

is commonly referred to as a linear probe (Alain and

Bengio, 2017). We argue that a supervised model that

is trained on annotated data from the target domain

gives a reasonable upper bound on the achievable

segmentation performance. The total domain gap

between the target domain T and the source domain

S then is the difference between the performance of

this model and a model trained on source data only.

Again, we retrain the linear classiﬁcation layer of this

source model with data from the target domain. This

source model achieves 39.6% mIoU

compared to

the upper bound of 68.0% mIoU

. Thus, we can

conclude that the total domain gap between S and T

is equal to 28.4% mIoU

points. In comparison, our

method achieves 59.0% mIoU

when retraining the

classiﬁcation layer. This reduces our estimate of the

remaining domain gap to ≈ 9%. This equals a reduc-

tion of the domain gap by 68.3% when compared to

the model that is trained on source data only.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

138

Figure 2: Examples of translated images. The left column shows synthetic images from the GTA5 dataset. The right column

shows the corresponding translated images.

5 CONCLUSION

In this work, we have investigated the problem of

unsupervised domain adaptation for semantic seg-

mentation. We proposed two complementary ap-

proaches in order to reduce the gap between the

data of a synthetic source domain and the real-world

target domain. More speciﬁcally, we have shown

that an adversarial image-to-image translation model

that is trained with an auxiliary segmentation task

on images of both domains yields signiﬁcantly better

results. We show that pseudo-labels can be leveraged

to improve this process. The combination of the

proposed methods outperforms previous state-of-the-

art combinations of image-to-image translation and

semi-supervised learning for domain adaptation on

relevant benchmarks by a considerable margin.

REFERENCES

Alain, G. and Bengio, Y. (2017). Understanding intermedi-

ate layers using linear classiﬁer probes. In 5th Interna-

tional Conference on Learning Representations, ICLR

2017, Toulon, France, April 24-26, 2017, Workshop

Track Proceedings. OpenReview.net. 8

Chen, L., Papandreou, G., Kokkinos, I., Murphy, K., and

Yuille, A. L. (2018). Deeplab: Semantic image

segmentation with deep convolutional nets, atrous

convolution, and fully connected crfs. IEEE Trans.

Pattern Anal. Mach. Intell., 40(4):834–848. 2, 7

Chen, Y., Chen, W., Chen, Y., Tsai, B., Wang, Y. F., and

Sun, M. (2017). No more discrimination: Cross city

adaptation of road scene segmenters. In IEEE Inter-

national Conference on Computer Vision, ICCV 2017,

Venice, Italy, October 22-29, 2017, pages 2011–2020.

IEEE Computer Society. 2

Chen, Y., Lin, Y., Yang, M., and Huang, J. (2019). Crdoco:

Pixel-level domain transfer with cross-domain consis-

tency. In IEEE Conference on Computer Vision and

Pattern Recognition, CVPR 2019, Long Beach, CA,

USA, June 16-20, 2019, pages 1791–1800. Computer

Vision Foundation / IEEE. 2

Choi, J., Kim, T., and Kim, C. (2019). Self-ensembling

with gan-based data augmentation for domain adap-

tation in semantic segmentation. In 2019 IEEE/CVF

International Conference on Computer Vision, ICCV

2019, Seoul, Korea (South), October 27 - November

2, 2019, pages 6829–6839. IEEE. 2, 6, 7, 8

Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler,

M., Benenson, R., Franke, U., Roth, S., and Schiele,

B. (2016). The cityscapes dataset for semantic ur-

ban scene understanding. In 2016 IEEE Conference

on Computer Vision and Pattern Recognition, CVPR

2016, Las Vegas, NV, USA, June 27-30, 2016, pages

3213–3223. IEEE Computer Society. 6

Deng, J., Dong, W., Socher, R., Li, L., Li, K., and Fei-Fei,

L. (2009). Imagenet: A large-scale hierarchical image

database. In 2009 IEEE Computer Society Conference

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation

139

on Computer Vision and Pattern Recognition (CVPR

2009), 20-25 June 2009, Miami, Florida, USA, pages

248–255. IEEE Computer Society. 3, 7

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Courville, A. C., and

Bengio, Y. (2014). Generative adversarial nets. In

Ghahramani, Z., Welling, M., Cortes, C., Lawrence,

N. D., and Weinberger, K. Q., editors, Advances in

Neural Information Processing Systems 27: Annual

Conference on Neural Information Processing Sys-

tems 2014, December 8-13 2014, Montreal, Quebec,

Canada, pages 2672–2680. 4, 7

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep

residual learning for image recognition. In 2016

IEEE Conference on Computer Vision and Pattern

Recognition, CVPR 2016, Las Vegas, NV, USA, June

27-30, 2016, pages 770–778. IEEE Computer Society.

Hoffman, J., Tzeng, E., Park, T., Zhu, J., Isola, P., Saenko,

K., Efros, A. A., and Darrell, T. (2018). Cycada:

Cycle-consistent adversarial domain adaptation. In

Dy, J. G. and Krause, A., editors, Proceedings of the

35th International Conference on Machine Learning,

ICML 2018, Stockholmsm

assan, Stockholm, Sweden,

July 10-15, 2018, volume 80 of Proceedings of Ma-

chine Learning Research, pages 1994–2003. PMLR.

2, 6

Hoffman, J., Wang, D., Yu, F., and Darrell, T. (2016). Fcns

in the wild: Pixel-level adversarial and constraint-

based adaptation. CoRR, abs/1612.02649. 2

Hong, W., Wang, Z., Yang, M., and Yuan, J. (2018).

Conditional generative adversarial network for struc-

tured domain adaptation. In 2018 IEEE Confer-

ence on Computer Vision and Pattern Recognition,

CVPR 2018, Salt Lake City, UT, USA, June 18-22,

2018, pages 1335–1344. Computer Vision Foundation

/ IEEE Computer Society. 2

Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2018).

Progressive growing of gans for improved quality, sta-

bility, and variation. In 6th International Conference

on Learning Representations, ICLR 2018, Vancouver,

BC, Canada, April 30 - May 3, 2018, Conference

Track Proceedings. OpenReview.net. 3

Ke, Z., Wang, D., Yan, Q., Ren, J. S. J., and Lau, R.

W. H. (2019). Dual student: Breaking the limits

of the teacher in semi-supervised learning. In 2019

IEEE/CVF International Conference on Computer Vi-

sion, ICCV 2019, Seoul, Korea (South), October 27 -

November 2, 2019, pages 6727–6735. IEEE. 2

Kingma, D. P. and Ba, J. (2015). Adam: A method for

stochastic optimization. In Bengio, Y. and LeCun,

Y., editors, 3rd International Conference on Learning

Representations, ICLR 2015, San Diego, CA, USA,

May 7-9, 2015, Conference Track Proceedings. 7

Lee, C., Batra, T., Baig, M. H., and Ulbricht, D. (2019).

Sliced wasserstein discrepancy for unsupervised do-

main adaptation. In IEEE Conference on Computer

Vision and Pattern Recognition, CVPR 2019, Long

Beach, CA, USA, June 16-20, 2019, pages 10285–

10295. Computer Vision Foundation / IEEE. 6

Lee, D.-H. (2013). Pseudo-label: The simple and efﬁcient

semi-supervised learning method for deep neural net-

works. In Workshop on challenges in representation

learning, ICML, volume 3, page 896. 2

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully

convolutional networks for semantic segmentation. In

IEEE Conference on Computer Vision and Pattern

Recognition, CVPR 2015, Boston, MA, USA, June 7-

12, 2015, pages 3431–3440. IEEE Computer Society.

Mao, X., Li, Q., Xie, H., Lau, R. Y. K., Wang, Z., and Smol-

ley, S. P. (2017). Least squares generative adversarial

networks. In IEEE International Conference on Com-

puter Vision, ICCV 2017, Venice, Italy, October 22-29,

2017, pages 2813–2821. IEEE Computer Society. 7

Murez, Z., Kolouri, S., Kriegman, D. J., Ramamoorthi,

R., and Kim, K. (2018). Image to image translation

for domain adaptation. In 2018 IEEE Conference

on Computer Vision and Pattern Recognition, CVPR

2018, Salt Lake City, UT, USA, June 18-22, 2018,

pages 4500–4509. Computer Vision Foundation /

IEEE Computer Society. 2

Odena, A., Olah, C., and Shlens, J. (2017). Conditional

image synthesis with auxiliary classiﬁer gans. In

Precup, D. and Teh, Y. W., editors, Proceedings of the

34th International Conference on Machine Learning,

ICML 2017, Sydney, NSW, Australia, 6-11 August

2017, volume 70 of Proceedings of Machine Learning

Research, pages 2642–2651. PMLR. 4

Pizzati, F., de Charette, R., Zaccaria, M., and Cerri, P.

(2020). Domain bridge for unpaired image-to-image

translation and unsupervised domain adaptation. In

IEEE Winter Conference on Applications of Computer

Vision, WACV 2020, Snowmass Village, CO, USA,

March 1-5, 2020, pages 2979–2987. IEEE. 2

Richter, S. R., Vineet, V., Roth, S., and Koltun, V. (2016).

Playing for data: Ground truth from computer games.

In Leibe, B., Matas, J., Sebe, N., and Welling, M., ed-

itors, Computer Vision - ECCV 2016 - 14th European

Conference, Amsterdam, The Netherlands, October

11-14, 2016, Proceedings, Part II, volume 9906 of

Lecture Notes in Computer Science, pages 102–118.

Springer. 6

Ros, G., Sellart, L., Materzynska, J., V

azquez, D., and

opez, A. M. (2016). The SYNTHIA dataset: A large

collection of synthetic images for semantic segmen-

tation of urban scenes. In 2016 IEEE Conference

on Computer Vision and Pattern Recognition, CVPR

2016, Las Vegas, NV, USA, June 27-30, 2016, pages

3234–3243. IEEE Computer Society. 6

Sankaranarayanan, S., Balaji, Y., Jain, A., Lim, S., and

Chellappa, R. (2018). Learning from synthetic data:

Addressing domain shift for semantic segmentation.

In 2018 IEEE Conference on Computer Vision and

Pattern Recognition, CVPR 2018, Salt Lake City, UT,

USA, June 18-22, 2018, pages 3752–3761. Computer

Vision Foundation / IEEE Computer Society. 2

Simonyan, K. and Zisserman, A. (2015). Very deep convo-

lutional networks for large-scale image recognition. In

Bengio, Y. and LeCun, Y., editors, 3rd International

Conference on Learning Representations, ICLR 2015,

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

140

San Diego, CA, USA, May 7-9, 2015, Conference

Track Proceedings. 3, 6

Taigman, Y., Polyak, A., and Wolf, L. (2017). Unsupervised

cross-domain image generation. In 5th International

Conference on Learning Representations, ICLR 2017,

Toulon, France, April 24-26, 2017, Conference Track

Proceedings. OpenReview.net. 3

Tarvainen, A. and Valpola, H. (2017). Mean teachers

are better role models: Weight-averaged consistency

targets improve semi-supervised deep learning results.

In Guyon, I., von Luxburg, U., Bengio, S., Wallach,

H. M., Fergus, R., Vishwanathan, S. V. N., and

Garnett, R., editors, Advances in Neural Information

Processing Systems 30: Annual Conference on Neural

Information Processing Systems 2017, December 4-9,

2017, Long Beach, CA, USA, pages 1195–1204. 2, 5

Toldo, M., Maracani, A., Michieli, U., and Zanuttigh, P.

(2020). Unsupervised domain adaptation in semantic

segmentation: a review. CoRR, abs/2005.10876. 2

van Engelen, J. E. and Hoos, H. H. (2020). A survey on

semi-supervised learning. Mach. Learn., 109(2):373–

440. 2

Wang, H., Shen, T., Zhang, W., Duan, L., and Mei, T.

(2020a). Classes matter: A ﬁne-grained adversarial

approach to cross-domain semantic segmentation. In

Vedaldi, A., Bischof, H., Brox, T., and Frahm, J.,

editors, Computer Vision - ECCV 2020 - 16th Euro-

pean Conference, Glasgow, UK, August 23-28, 2020,

Proceedings, Part XIV, volume 12359 of Lecture

Notes in Computer Science, pages 642–659. Springer.

Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao,

Y., Liu, D., Mu, Y., Tan, M., Wang, X., Liu, W., and

Xiao, B. (2021). Deep high-resolution representation

learning for visual recognition. IEEE Trans. Pattern

Anal. Mach. Intell., 43(10):3349–3364. 2

Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., and Bailey, J.

(2019). Symmetric cross entropy for robust learning

with noisy labels. In 2019 IEEE/CVF International

Conference on Computer Vision, ICCV 2019, Seoul,

Korea (South), October 27 - November 2, 2019, pages

322–330. IEEE. 4

Wang, Z., Yu, M., Wei, Y., Feris, R., Xiong, J., Hwu,

W., Huang, T. S., and Shi, H. (2020b). Differential

treatment for stuff and things: A simple unsupervised

domain adaptation method for semantic segmentation.

In 2020 IEEE/CVF Conference on Computer Vision

and Pattern Recognition, CVPR 2020, Seattle, WA,

USA, June 13-19, 2020, pages 12632–12641. Com-

puter Vision Foundation / IEEE. 6

Wu, Z., Han, X., Lin, Y., Uzunbas, M. G., Goldstein,

T., Lim, S., and Davis, L. S. (2018). DCAN: dual

channel-wise alignment networks for unsupervised

scene adaptation. In Ferrari, V., Hebert, M., Smin-

chisescu, C., and Weiss, Y., editors, Computer Vision

- ECCV 2018 - 15th European Conference, Munich,

Germany, September 8-14, 2018, Proceedings, Part V,

volume 11209 of Lecture Notes in Computer Science,

pages 535–552. Springer. 6

Xie, Q., Luong, M., Hovy, E. H., and Le, Q. V. (2020).

Self-training with noisy student improves imagenet

classiﬁcation. In 2020 IEEE/CVF Conference on

Computer Vision and Pattern Recognition, CVPR

2020, Seattle, WA, USA, June 13-19, 2020, pages

10684–10695. Computer Vision Foundation / IEEE.

Yang, Y., Lao, D., Sundaramoorthi, G., and Soatto, S.

(2020). Phase consistent ecological domain adapta-

tion. In 2020 IEEE/CVF Conference on Computer

Vision and Pattern Recognition, CVPR 2020, Seattle,

WA, USA, June 13-19, 2020, pages 9008–9017. Com-

puter Vision Foundation / IEEE. 6

Yu, F. and Koltun, V. (2016). Multi-scale context aggrega-

tion by dilated convolutions. In Bengio, Y. and LeCun,

Y., editors, 4th International Conference on Learning

Representations, ICLR 2016, San Juan, Puerto Rico,

May 2-4, 2016, Conference Track Proceedings. 2

Zhang, Y., David, P., and Gong, B. (2017). Curriculum

domain adaptation for semantic segmentation of urban

scenes. In IEEE International Conference on Com-

puter Vision, ICCV 2017, Venice, Italy, October 22-29,

2017, pages 2039–2049. IEEE Computer Society. 7

Zhou, Q., Feng, Z., Cheng, G., Tan, X., Shi, J., and Ma,

L. (2020). Uncertainty-aware consistency regulariza-

tion for cross-domain semantic segmentation. CoRR,

abs/2004.08878. 5, 6, 7, 8

Zhu, J., Park, T., Isola, P., and Efros, A. A. (2017). Unpaired

image-to-image translation using cycle-consistent ad-

versarial networks. In IEEE International Conference

on Computer Vision, ICCV 2017, Venice, Italy, Octo-

ber 22-29, 2017, pages 2242–2251. IEEE Computer

Society. 2

Zou, Y., Yu, Z., Kumar, B. V. K. V., and Wang, J.

(2018). Unsupervised domain adaptation for seman-

tic segmentation via class-balanced self-training. In

Ferrari, V., Hebert, M., Sminchisescu, C., and Weiss,

Y., editors, Computer Vision - ECCV 2018 - 15th

European Conference, Munich, Germany, September

8-14, 2018, Proceedings, Part III, volume 11207 of

Lecture Notes in Computer Science, pages 297–313.

Springer. 6

Semantically Consistent Image-to-Image Translation for Unsupervised Domain Adaptation

141