IFMix: Utilizing Intermediate Filtered Images for Domain Adaptation in

Classiﬁcation

Saeed Bakhshi Germi

and Esa Rahtu

Computer Vision Group, Tampere University, Tampere, Finland

ﬁ

Keywords:

Domain Adaptation, Filtered Images, Classiﬁcation, Mixup Technique.

Abstract:

This paper proposes an iterative intermediate domain generation method using low- and high-pass ﬁlters.

Domain shift is one of the prime reasons for the poor generalization of trained models in most real-life appli-

cations. In a typical case, the target domain differs from the source domain due to either controllable factors

(e.g., different sensors) or uncontrollable factors (e.g., weather conditions). Domain adaptation methods bridge

this gap by training a domain-invariant network. However, a signiﬁcant gap between the source and the target

domains would still result in bad performance. Gradual domain adaptation methods utilize intermediate do-

mains that gradually shift from the source to the target domain to counter the effect of the signiﬁcant gap. Still,

the assumption of having sufﬁciently large intermediate domains at hand for any given task is hard to fulﬁll in

real-life scenarios. The proposed method utilizes low- and high-pass ﬁlters to create two distinct representa-

tions of a single sample. After that, the ﬁltered samples from two domains are mixed with a dynamic ratio to

create intermediate domains, which are used to train two separate models in parallel. The ﬁnal output is ob-

tained by averaging out both models. The method’s effectiveness is demonstrated with extensive experiments

on public benchmark datasets: Ofﬁce-31, Ofﬁce-Home, and VisDa-2017. The empirical evaluation suggests

that the proposed method performs better than the current state-of-the-art works.

1 INTRODUCTION

With the increasing popularity of deep learning algo-

rithms in the heavy machine industry and the inclu-

sion of artiﬁcial intelligence in new regulations (e.g.,

EU AI Act) and safety standards (e.g., ISO/IEC JTC

1/SC 42 Committee), the practical issues of utilizing

such algorithms in safety-critical applications have

become more apparent. One of the challenges for

any practical application of a deep learning algorithm

is collecting and labeling a large dataset for training

the algorithm while considering the safety criteria for

the application (Bakhshi Germi and Rahtu, 2022b).

A standard method to deal with this issue is utiliz-

ing transfer learning (Zhuang et al., 2021), where the

model is trained with a label-rich source dataset (e.g.,

synthesized or simulated data) and ﬁne-tuned on a

much smaller target dataset (e.g., data collected from

the real world). However, a signiﬁcant gap between

these two domains would result in poor performance.

Gradual domain adaptation (GDA) deals with the

gap problem by adding data from intermediate do-

mains that interpolate between the source and the tar-

https://orcid.org/0000-0003-3048-220X

https://orcid.org/0000-0001-8767-0864

get domains (Kumar et al., 2020). The intermedi-

ate domains are assumed to be available with sufﬁ-

cient data for the training process. The accuracy of

GDA methods is highly dependent on the distance be-

tween the source and the target domains. Moreover,

GDA methods are usually unsupervised and do not

require labels from intermediate or target domains.

While unsupervised methods attract more attention in

the research community, using a small labeled sub-

set from the target domain is more realistic in real-

world applications. Various annotation tools (Ad-

hikari and Huttunen, 2021) and denoising techniques

(Bakhshi Germi and Rahtu, 2022a) could be utilized

to help with gathering the required labeled subset.

Meanwhile, intermediate domains do not naturally

exist for most real-world applications. Thus, this pa-

per focuses on generating intermediate domains based

on a large labeled source dataset and a small labeled

target dataset.

This paper proposes IFMix, a domain adaptation

algorithm that utilizes a ﬁltered-image-based mixup

technique to create intermediate domains iteratively.

A new domain is created by merging the low-pass or

high-pass ﬁltered images from both domains with a

dynamic ratio. The images are chosen from the same

Germi, S. and Rahtu, E.

IFMix: Utilizing Intermediate Filtered Images for Domain Adaptation in Classiﬁcation.

DOI: 10.5220/0011713600003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

205-211

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

205

Source Domain

Target Domain

Mixup Unit

Low-passed sample

High-Passed sample

CNN

Cross-entropy Loss

Co-convergence Term

Figure 1: The overall structure of the proposed method. Two samples of the same category are chosen from two domains to

be mixed. The mixup unit utilizes low-pass and high-pass ﬁlters to mix images with different ratios. The resulting images

are used as training samples for two separate models. Each model is trained with a categorical cross-entropy loss. A co-

convergence term is utilized to ensure the convergence of both models towards the same point.

category in both domains to keep the labels intact. Af-

ter that, the proposed method utilizes the intermedi-

ate domains to train two separate models in parallel.

Both models’ average output is considered the pro-

posed method’s ﬁnal output. The intuition behind the

proposed method is that a supervised method that re-

lies on a small amount of data from the target domain

would be practical and realistic, the iterative domain

creation would compensate for the lack of data in real-

world applications, and the two models develop dif-

ferent perspectives based on their respective ﬁlters.

The main difference between the proposed

method and previous works is utilizing a small la-

beled target dataset to create intermediate domains,

resulting in accurate labels instead of pseudo-labels.

Also, using the low- and high-pass ﬁlters would re-

sult in two distinct representations of the same sam-

ple, creating substantially different intermediate do-

mains for training two different models. Moreover,

the iterative and gradual nature of the algorithm en-

sures that the model is not overwhelmed by new in-

formation while the gap between the two domains is

breached. The effectiveness of the proposed method

is shown by comparing the performance with previ-

ous state-of-the-art methods in standard public bench-

marks such as Ofﬁce-31 (Saenko et al., 2010), Ofﬁce-

Home (Venkateswara et al., 2017), and VisDa-2017

(Peng et al., 2017). The main contributions of this

paper are summarized as follows:

• Proposing an iterative intermediate domain cre-

ation technique based on ﬁltered images to bridge

the gap between the source and the target do-

mains.

• Providing a practical domain adaptation algorithm

based on the proposed intermediate domains.

• Providing empirical evaluation with extensive ex-

periments on three standard benchmarks to show

the effectiveness of the proposed method.

The rest of the paper is structured as follows. Sec-

tion 2 covers the related works. Next, Section 3 ex-

plains the proposed method in detail. After that, Sec-

tion 4 deals with the experiments and the empirical

evaluation to show the effectiveness of the proposed

method. Finally, Section 5 concludes the work.

2 RELATED WORKS

2.1 Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) methods uti-

lize domain-invariant representation to generalize a

model from a rich-labeled source domain to an un-

labeled target domain (Wilson and Cook, 2020). The

process can be done by either optimizing distribution

discrepancy metrics (e.g., maximum mean discrep-

ancy) (Li et al., 2021a; Peng et al., 2019) or utilizing

adversarial training (Li et al., 2021b; Liu et al., 2019;

Wang et al., 2019). On top of that, utilizing pseudo-

labeling ideas from semi-supervised learning meth-

ods improves the performance of UDA algorithms

(Chen et al., 2020; Liang et al., 2020; Liang et al.,

2021; Liu et al., 2021; Zhang et al., 2021b). More-

over, the natural advantage of transformers in ex-

tracting transferable representations was studied fur-

ther for application in domain adaptation (Ma et al.,

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

206

Source Domain Target Domain

Figure 2: The creation of multiple intermediate domains by the proposed method. The shown samples are not ﬁltered to

understand better how the method works. Samples progress from the source domain (left) to the target domain (right) with

each iteration based on the value of H.

2021; Xu et al., 2021; Yang et al., 2021a). Unsu-

pervised methods have been the research focus for a

while in academic applications. However, utilizing

a small labeled dataset could result in a performance

surge without signiﬁcantly increasing the overall cost

of gathering data.

2.2 Gradual Domain Adaptation

Gradual domain adaptation methods utilize interme-

diate domains to improve the performance of basic

domain adaptation techniques (Choi et al., 2020; Cui

et al., 2020; Dai et al., 2021; Hsu et al., 2020). GDA

methods utilize generative models (e.g., generative

adversarial networks) to create an intermediate do-

main by mixing the source and the target data at an

arbitrary ratio (Sagawa and Hino, 2022). By doing so,

the model can learn common features shared between

two domains. In the original work, Kumar assumed

that the intermediate domains gradually shift from

the source to the target domain, and their sequence

is known prior to learning (Kumar et al., 2020). How-

ever, the method is effective even if the sequence of

these domains is unknown (Chen and Chao, 2021;

Zhang et al., 2021a) or when no intermediate domain

is available (Abnar et al., 2021; Na et al., 2021b). The

main difference between current state-of-the-art GDA

algorithms is their technique for creating intermediate

domains.

2.3 Mixup Technique

Mixup techniques are a family of data augmentation

methods based on mixing two or more data points.

Mixup and its variants have proven helpful in super-

vised and semi-supervised learning (Berthelot et al.,

2019; Yun et al., 2019; Zhang et al., 2017). Some

recent domain adaptation methods tried utilizing this

technique to create a continuous latent space across

domains (Wu et al., 2020; Xu et al., 2020), obtain

pseudo labels for intermediate domains (Na et al.,

2021b; Yan et al., 2020; Yang et al., 2021b), or gener-

ate more positive/negative samples (Kalantidis et al.,

2020; Zhang et al., 2022; Zhu et al., 2021).

This paper utilizes the intermediate domains from

GDA, a mixup technique based on low- and high-

pass ﬁlters, and a small labeled subset from the tar-

get domain to achieve high performance in real-world

scenarios. The assumptions in this paper are tai-

lored around practical use cases of domain adapta-

tion where a large labeled source domain and a small

labeled target domain are available. While similar

works exist in this ﬁeld, the proposed method outper-

forms the existing state-of-the-art, as shown in Sec-

tion 4.

3 PROPOSED METHOD

This section presents the details of the proposed

method, as shown in Figure 1. Let D

= {(x

, y

)}

i=1

be the labeled dataset from the source domain, D

{(x

)}

j=1

be the unlabeled dataset from the target

domain, and D

= {(x

, y

)}

k=1

be the labeled sub-

set from the target domain. The task is transferring

knowledge from D

to D

when there is a large distri-

bution gap between them.

3.1 Iterative Filtered Mixup

The proposed method selects random samples with

the same category label from D

and D

, applies low-

and high-pass ﬁlters on them, and mixes them to cre-

ate new samples as follows:

= (1 − H)× LoPass(x

) +H × LoPass(x

)

= (1 − H)× HiPass(x

) +H × HiPass(x

)

(1)

Where (0 ≤ H ≤ 1) denotes a dynamic ratio for the

mixing step, LoPass and HiPass denote the low-pass

and high-pass ﬁlter functions, respectively. These ﬁl-

ters could be implemented using the Gaussian ﬁlter

function in the Multidimensional Image Processing

package (scipy.ndimage). Moreover, the labels y

and y

for generated samples would be the same as

the original label y

due to choosing samples from the

same category. Finally, the mixing ratio H is updated

based on the number of epochs as follows:

i+1

= H

+ α × t (2)

IFMix: Utilizing Intermediate Filtered Images for Domain Adaptation in Classiﬁcation

207

Where α is a positive constant and t is the current

number of epochs. Two labeled datasets, D

and D

are created with each iteration. These intermediate

datasets ﬁll the gap between the source and the target

domains, as shown in Figure 2. Note that the ﬁgure

shows unﬁltered samples for a more straightforward

interpretation of how the algorithm works.

3.2 Training and Loss Functions

In the next step, two models are trained on D

and

using the categorical cross-entropy loss function:

cce

∑

× log



y|x



cce

∑

× log



y|x



(3)

Where p(y|x

) and q(y|x

) denote the predicted class

for each network on their respective input, and B is the

batch size. The models are trained separately for a

few epochs (warm-up period) to ensure they gain dif-

ferent perspectives without the inﬂuence of the other

model.

3.3 Output and Co-Convergence Term

With each model training to recognize different char-

acteristics of a given sample, their average output is

used to determine the ﬁnal output of the algorithm.

Since the models should converge towards the same

goal, a co-convergence term is added to the overall

loss after the warm-up period. This term ensures that

each model can inﬂuence the other model slightly to

reach a similar conclusion on their output.

cct

∑

× log



y|x



+ q



y|x



(4)

3.4 Overall Process

The overall process of the IFMix algorithm is sum-

marized in Algorithm 1. The algorithm starts with

creating the intermediate domains in each iteration.

Then two networks are trained with the new interme-

diate domains using the deﬁned loss functions. The

co-convergence term is added after the warm-up pe-

riod to allow the models to develop unique character-

istics without the inﬂuence of the other model.

In experiments, the mixup ratio H is updated every

few epochs to prevent potential divergence of models.

Algorithm 1: IFMix Algorithm.

Require: Source dataset D

, Labeled Target

subset D

, Number of epochs T , Batch size

B, Warm-up period W , Mixup ratio H,

Mixup increment rate α

1: for t ∈ 1, . . . , T do

2: Select samples from same category in D

and D

3: Create intermediate domains D

and D

using Eq. 1

4: for b ∈ 1, . . . , B do

5: Update loss functions L

cce

and L

cce

using Eq. 3

6: if i ≥ W then

7: Update co-convergence term L

cct

using Eq. 4

8: end if

9: end for

10: Update the mixup ratio

using Eq. 2

11: end for

4 EXPERIMENTS &

EVALUATION

To evaluate the proposed method, three different do-

main adaptation benchmarks are chosen so that the

performance of the proposed method can be com-

pared with state-of-the-art methods. In each exper-

iment, 5% of samples from the target domain are

selected as labeled target subsets for the proposed

method, and the remaining 95% of samples are left

as test data.

4.1 Ofﬁce-31

Ofﬁce-31 (Saenko et al., 2010), a domain adaptation

benchmark, provides samples for 31 categories from

three domains. These domains are denoted as A for

images taken from Amazon.com, D for images taken

with a DSLR camera, and W for images taken with a

webcam. The dataset has around 4000 samples, mak-

ing it a perfect benchmark for proof of concept.

4.2 Ofﬁce-Home

Ofﬁce-Home (Venkateswara et al., 2017), a domain

adaptation benchmark, provides samples for 65 cate-

gories from four domains. These domains are denoted

as A for arts and paintings, C for clipart images, P for

product images without a background, and R for real-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

208

Table 1: Accuracy (%) on the Ofﬁce-31 dataset. The best accuracy is indicated in bold, and the second best is underlined.

Method A → D A → W D → A D → W W → A W → D Average

GSDA (Hu et al., 2020) 94.8 95.7 73.5 99.1 74.9 100 89.7

SRDC (Tang et al., 2020) 95.8 95.7 76.7 99.2 77.1 100 90.8

RSDA (Gu et al., 2020) 95.8 96.1 77.4 99.3 78.9 100 91.1

FixBi (Na et al., 2021b) 95 96.1 78.7 99.3 79.4 100 91.4

CoVi (Na et al., 2021a) 98 97.6 77.5 99.3 78.4 100 91.8

IFMix (Ours) 97.6 97.5 77.9 99.3 79.7 100 92

Table 2: Accuracy (%) on the Ofﬁce-Home dataset. The best accuracy is indicated in bold, and the second best is underlined.

Method A → C A → P A → R C → A C → P C → R P → A P → C P → R R → A R → C R → P Average

MetaAlign (Wei et al., 2021) 59.3 76 80.2 65.7 74.7 75.1 65.7 56.5 81.6 74.1 61.1 85.2 71.3

FixBi (Na et al., 2021b) 58.1 77.3 80.4 67.7 79.5 78.1 65.8 57.9 81.7 76.4 62.9 86.7 72.7

CoVi (Na et al., 2021a) 58.5 78.1 80 68.1 80 77 66.4 60.2 82.1 76.6 63.6 86.5 73.1

CDTrans (Xu et al., 2021) 60.6 79.5 82.4 75.6 81 82.3 72.5 56.7 84.4 77 59.1 85.5 74.7

WinTR (Ma et al., 2021) 65.3 84.1 85 76.8 84.5 84.4 73.4 60 85.7 77.2 63.1 86.8 77.2

IFMix (Ours) 66.1 84 86.6 77.4 84.1 86.1 75.2 61.1 86.5 78.4 62.8 87.4 78

Table 3: Accuracy (%) on the VisDa-2017 dataset. The best accuracy is indicated in bold, and the second best is underlined.

Method Plane Bike Bus Car Horse Knife Motor Human Plant Skate Train Truck Average

CAN (Kang et al., 2019) 97 87.2 82.5 74.3 97.8 96.2 90.8 80.7 96.6 96.3 87.5 59.9 87.2

FixBi (Na et al., 2021b) 96.1 87.8 90.5 90.3 96.8 95.3 92.8 88.7 97.2 94.2 90.9 25.7 87.2

CDTrans (Xu et al., 2021) 97.1 90.5 82.4 77.5 96.6 96.1 93.6 88.6 97.9 86.9 90.3 62.8 88.4

CoVi (Na et al., 2021a) 96.8 85.6 88.9 88.6 97.8 93.4 91.9 87.6 96 93.8 93.6 48.1 88.5

WinTR (Ma et al., 2021) 98.7 91.2 93 91.9 98.1 96.1 94 72.7 97 95.5 95.3 57.9 90.1

IFMix (Ours) 98.2 91.7 92.9 92.2 98.5 96.5 93.7 88 98 95.5 94.8 61.8 91.8

world images taken with a camera. The dataset has

around 15000 samples, making it a more challenging

task than Ofﬁce-31.

4.3 VisDa-2017

VisDa-2017 (Peng et al., 2017), a domain adaptation

benchmark, provides samples for 12 categories from

two domains, simulated and real-world. The dataset

has around 280000 samples, making it a complex and

realistic benchmark for domain adaptation problems.

4.4 Hyper-Parameters

In the experiments with Ofﬁce datasets, ResNet-50

with stochastic gradient descent (SGD) is used as the

base model. The initial learning rate is 0.001, the mo-

mentum is 0.9, the weight decay is 0.005, the initial

mixup ratio is 0.05 with a 0.05 increment every 10

epochs, and the total number of epochs is 200. In

the experiments with VisDA dataset, the base model

is swapped to ResNet-101. The initial learning rate

is 0.0001, the initial mixup ratio is 0.1 with a 0.1

increment every 5 epochs, and the total number of

epochs is 50. In all experiments, the models utilize

pre-trained weights on ImageNet (Russakovsky et al.,

2015).

4.5 Results and Comparison

Table 1 holds the results for the Ofﬁce-31 dataset.

Six different tasks are experimented upon, and the

results are compared with state-of-the-art methods.

The accuracy of state-of-the-art methods is obtained

from their respective published papers. The results

from each task indicate that the proposed method is

competitive. The average accuracy of the proposed

method is 92%, which is a slight improvement over

the previous best method, CoVi (Na et al., 2021a).

As stated before, the Ofﬁce-31 dataset was utilized

to prove that the proposed method works as intended,

even if the improvement is slight and negligible.

Table 2 holds the results for the Ofﬁce-Home

dataset. Twelve different tasks are experimented

upon, and the results are compared with state-of-the-

art methods. Similar to previous experiments, the ac-

curacy of state-of-the-art methods is obtained from

their respective published papers. The results from

each task indicate that the proposed method is still

competitive. The average accuracy of the proposed

method is 78%, which is an improvement over the

previous best method, WinTR (Ma et al., 2021). Note

that the proposed method outperformed CoVi (Na

et al., 2021a), the previous best method on the Ofﬁce-

31 dataset, by 4.9% on average. This experiment of-

fers more insight into the value of utilizing the pro-

posed method. While the proposed method slightly

outperforms the alternatives in this case, it also of-

fers a more robust solution that works on different

datasets.

Table 3 holds the results for the VisDa-2017

dataset. The results are compared on category and

overall level. Similar to previous experiments, the ac-

curacy of state-of-the-art methods is obtained from

their respective published papers. The results from

each category indicate that the proposed method is

IFMix: Utilizing Intermediate Filtered Images for Domain Adaptation in Classiﬁcation

209

operating as intended. The average accuracy of the

proposed method is 91.8%, which is a signiﬁcant im-

provement over the previous best method, WinTR

(Ma et al., 2021). The proposed method offers a no-

ticeable improvement in this experiment.

5 CONCLUSION

This paper proposed a practical domain adaptation

method that utilizes a labeled subset from the target

domain and low- and high-pass ﬁlters to create inter-

mediate domains. The iterative creation of interme-

diate domains helps the model quickly adapt despite

a signiﬁcant gap between domains. The effectiveness

of the proposed method is shown with empirical ex-

periments on public benchmark datasets. The pro-

posed method outperforms the current state-of-the-art

methods by a noticeable margin while maintaining ro-

bustness over different datasets.

ACKNOWLEDGEMENT

This research is part of a Ph.D. study co-funded by

Tampere University and Forum for Intelligent Ma-

chines ry (FIMA).

REFERENCES

Abnar, S., Berg, R. v. d., Ghiasi, G., Dehghani, M., Kalch-

brenner, N., and Sedghi, H. (2021). Gradual domain

adaptation in the wild: When intermediate distribu-

tions are absent.

Adhikari, B. and Huttunen, H. (2021). Iterative bounding

box annotation for object detection. In International

Conference on Pattern Recognition (ICPR).

Bakhshi Germi, S. and Rahtu, E. (2022a). Enhanced data-

recalibration: Utilizing validation data to mitigate

instance-dependent noise in classiﬁcation. In Image

Analysis and Processing (ICIAP).

Bakhshi Germi, S. and Rahtu, E. (2022b). A practical

overview of safety concerns and mitigation methods

for visual deep learning algorithms. In Proceed-

ings of the Workshop on Artiﬁcial Intelligence Safety

(SafeAI).

Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N.,

Oliver, A., and Raffel, C. A. (2019). Mixmatch: A

holistic approach to semi-supervised learning. In Ad-

vances in Neural Information Processing Systems.

Chen, H.-Y. and Chao, W.-L. (2021). Gradual domain adap-

tation without indexed intermediate domains. In Ad-

vances in Neural Information Processing Systems.

Chen, M., Zhao, S., Liu, H., and Cai, D. (2020).

Adversarial-learned loss for domain adaptation. Pro-

ceedings of the AAAI Conference on Artiﬁcial Intelli-

gence, 34(4).

Choi, J., Choi, Y., Kim, J., Chang, J., Kwon, I., Gwon,

Y., and Min, S. (2020). Visual domain adaptation

by consensus-based transfer to intermediate domain.

Proceedings of the AAAI Conference on Artiﬁcial In-

telligence, 34(7).

Cui, S., Wang, S., Zhuo, J., Su, C., Huang, Q., and Tian,

Q. (2020). Gradually vanishing bridge for adversarial

domain adaptation. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Dai, Y., Liu, J., Sun, Y., Tong, Z., Zhang, C., and Duan,

L.-Y. (2021). Idm: An intermediate domain module

for domain adaptive person re-id. In Proceedings of

the IEEE/CVF International Conference on Computer

Vision (ICCV).

Gu, X., Sun, J., and Xu, Z. (2020). Spherical space domain

adaptation with robust pseudo-label loss. In Proceed-

ings of the IEEE/CVF Conference on Computer Vision

and Pattern Recognition (CVPR).

Hsu, H.-K., Yao, C.-H., Tsai, Y.-H., Hung, W.-C., Tseng,

H.-Y., Singh, M., and Yang, M.-H. (2020). Progres-

sive domain adaptation for object detection. In Pro-

ceedings of the IEEE/CVF Winter Conference on Ap-

plications of Computer Vision (WACV).

Hu, L., Kan, M., Shan, S., and Chen, X. (2020). Unsu-

pervised domain adaptation with hierarchical gradi-

ent synchronization. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Kalantidis, Y., Sariyildiz, M. B., Pion, N., Weinzaepfel, P.,

and Larlus, D. (2020). Hard negative mixing for con-

trastive learning. In Advances in Neural Information

Processing Systems.

Kang, G., Jiang, L., Yang, Y., and Hauptmann, A. G.

(2019). Contrastive adaptation network for unsu-

pervised domain adaptation. In Proceedings of the

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition (CVPR).

Kumar, A., Ma, T., and Liang, P. (2020). Understanding

self-training for gradual domain adaptation. In Pro-

ceedings of the 37th International Conference on Ma-

chine Learning (PMLR).

Li, S., Liu, C. H., Lin, Q., Wen, Q., Su, L., Huang, G., and

Ding, Z. (2021a). Deep residual correction network

for partial domain adaptation. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 43(7).

Li, S., Xie, M., Lv, F., Liu, C. H., Liang, J., Qin, C., and Li,

W. (2021b). Semantic concentration for domain adap-

tation. In Proceedings of the IEEE/CVF International

Conference on Computer Vision (ICCV).

Liang, J., Hu, D., and Feng, J. (2020). Do we really need to

access the source data? source hypothesis transfer for

unsupervised domain adaptation. In Proceedings of

the 37th International Conference on Machine Learn-

ing (PMLR).

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

210

Liang, J., Hu, D., and Feng, J. (2021). Domain adapta-

tion with auxiliary target domain-oriented classiﬁer.

In Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition (CVPR).

Liu, H., Long, M., Wang, J., and Jordan, M. (2019).

Transferable adversarial training: A general approach

to adapting deep classiﬁers. In Proceedings of the

36th International Conference on Machine Learning

(PMLR).

Liu, H., Wang, J., and Long, M. (2021). Cycle self-training

for domain adaptation. In Advances in Neural Infor-

mation Processing Systems.

Ma, W., Zhang, J., Li, S., Liu, C. H., Wang, Y., and Li, W.

(2021). Exploiting both domain-speciﬁc and invari-

ant knowledge via a win-win transformer for unsuper-

vised domain adaptation.

Na, J., Han, D., Chang, H. J., and Hwang, W. (2021a). Con-

trastive vicinal space for unsupervised domain adapta-

tion.

Na, J., Jung, H., Chang, H. J., and Hwang, W. (2021b).

Fixbi: Bridging domain spaces for unsupervised do-

main adaptation. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recog-

nition (CVPR).

Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., and

Wang, B. (2019). Moment matching for multi-source

domain adaptation. In Proceedings of the IEEE/CVF

International Conference on Computer Vision (ICCV).

Peng, X., Usman, B., Kaushik, N., Hoffman, J., Wang, D.,

and Saenko, K. (2017). Visda: The visual domain

adaptation challenge.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,

Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-

stein, M., and et al. (2015). Imagenet large scale vi-

sual recognition challenge. International Journal of

Computer Vision, 115(3).

Saenko, K., Kulis, B., Fritz, M., and Darrell, T. (2010).

Adapting visual category models to new domains. In

European Conference on Computer Vision (ECCV).

Sagawa, S. and Hino, H. (2022). Gradual domain adaptation

via normalizing ﬂows.

Tang, H., Chen, K., and Jia, K. (2020). Unsupervised do-

main adaptation via structurally regularized deep clus-

tering. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition (CVPR).

Venkateswara, H., Eusebio, J., Chakraborty, S., and Pan-

chanathan, S. (2017). Deep hashing network for un-

supervised domain adaptation. In IEEE Conference

on Computer Vision and Pattern Recognition (CVPR).

Wang, X., Li, L., Ye, W., Long, M., and Wang, J. (2019).

Transferable attention for domain adaptation. Pro-

ceedings of the AAAI Conference on Artiﬁcial Intel-

ligence, 33(1).

Wei, G., Lan, C., Zeng, W., and Chen, Z. (2021). Metaalign:

Coordinating domain alignment and classiﬁcation for

unsupervised domain adaptation. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition (CVPR).

Wilson, G. and Cook, D. J. (2020). A survey of unsuper-

vised deep domain adaptation. ACM Transactions on

Intelligent Systems and Technology, 11(5).

Wu, Y., Inkpen, D., and El-Roby, A. (2020). Dual mixup

regularized learning for adversarial domain adapta-

tion. In European Conference on Computer Vision

(ECCV).

Xu, M., Zhang, J., Ni, B., Li, T., Wang, C., Tian, Q., and

Zhang, W. (2020). Adversarial domain adaptation

with domain mixup. Proceedings of the AAAI Con-

ference on Artiﬁcial Intelligence, 34(4).

Xu, T., Chen, W., Wang, P., Wang, F., Li, H., and Jin, R.

(2021). Cdtrans: Cross-domain transformer for unsu-

pervised domain adaptation.

Yan, S., Song, H., Li, N., Zou, L., and Ren, L. (2020).

Improve unsupervised domain adaptation with mixup

training.

Yang, G., Tang, H., Zhong, Z., Ding, M., Shao, L., Sebe, N.,

and Ricci, E. (2021a). Transformer-based source-free

domain adaptation.

Yang, L., Wang, Y., Gao, M., Shrivastava, A., Weinberger,

K. Q., Chao, W.-L., and Lim, S.-N. (2021b). Deep co-

training with task decomposition for semi-supervised

domain adaptation. In Proceedings of the IEEE/CVF

International Conference on Computer Vision (ICCV).

Yun, S., Han, D., Oh, S. J., Chun, S., Choe, J., and Yoo,

Y. (2019). Cutmix: Regularization strategy to train

strong classiﬁers with localizable features. In Pro-

ceedings of the IEEE/CVF International Conference

on Computer Vision (ICCV).

Zhang, H., Cisse, M., Dauphin, Y. N., and Lopez-Paz, D.

(2017). Mixup: Beyond empirical risk minimization.

Zhang, Y., Deng, B., Jia, K., and Zhang, L. (2021a). Grad-

ual domain adaptation via self-training of auxiliary

models.

Zhang, Y., Li, J., and Wang, Z. (2022). Low-conﬁdence

samples matter for domain adaptation.

Zhang, Y., Wang, Z., and Mao, Y. (2021b). Rpn prototype

alignment for domain adaptive object detector. In Pro-

ceedings of the IEEE/CVF Conference on Computer

Vision and Pattern Recognition (CVPR).

Zhu, R., Zhao, B., Liu, J., Sun, Z., and Chen, C. W. (2021).

Improving contrastive learning by visualizing feature

transformation. In Proceedings of the IEEE/CVF In-

ternational Conference on Computer Vision (ICCV).

Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H.,

Xiong, H., and He, Q. (2021). A comprehensive sur-

vey on transfer learning. Proceedings of the IEEE,

109(1).

IFMix: Utilizing Intermediate Filtered Images for Domain Adaptation in Classiﬁcation

211