diff --git "a/09AyT4oBgHgl3EQfbffL/content/tmp_files/2301.00265v1.pdf.txt" "b/09AyT4oBgHgl3EQfbffL/content/tmp_files/2301.00265v1.pdf.txt"
new file mode 100644--- /dev/null
+++ "b/09AyT4oBgHgl3EQfbffL/content/tmp_files/2301.00265v1.pdf.txt"
@@ -0,0 +1,2716 @@
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+1
+Source-Free Unsupervised Domain Adaptation:
+A Survey
+Yuqi Fang, Pew-Thian Yap, Senior Member, IEEE, Weili Lin, Hongtu Zhu, and Mingxia Liu, Senior
+Member, IEEE
+Abstract—Unsupervised domain adaptation (UDA) via deep learning has attracted appealing attention for tackling domain-shift
+problems caused by distribution discrepancy across different domains. Existing UDA approaches highly depend on the accessibility of
+source domain data, which is usually limited in practical scenarios due to privacy protection, data storage and transmission cost, and
+computation burden. To tackle this issue, many source-free unsupervised domain adaptation (SFUDA) methods have been proposed
+recently, which perform knowledge transfer from a pre-trained source model to unlabeled target domain with source data inaccessible.
+A comprehensive review of these works on SFUDA is of great significance. In this paper, we provide a timely and systematic literature
+review of existing SFUDA approaches from a technical perspective. Specifically, we categorize current SFUDA studies into two groups,
+i.e., white-box SFUDA and black-box SFUDA, and further divide them into finer subcategories based on different learning strategies
+they use. We also investigate the challenges of methods in each subcategory, discuss the advantages/disadvantages of white-box and
+black-box SFUDA methods, conclude the commonly used benchmark datasets, and summarize the popular techniques for improved
+generalizability of models learned without using source data. We finally discuss several promising future directions in this field.
+Index Terms—Domain adaptation, source-free, unsupervised learning, survey.
+!
+1
+INTRODUCTION
+D
+EEP learning, based on deep neural networks with rep-
+resentation learning, has emerged as a promising tech-
+nique and made remarkable progress over the past decade,
+covering the field of computer vision [1], [2], medical data
+analysis [3], [4], natural language processing [5], [6], etc. For
+problems with multiple domains (e.g., different datasets or
+imaging sites), the typical learning process of a deep neural
+network is to transfer the model learned on a source domain
+to a target domain. However, performance degradation is
+often observed when there exists a distribution gap between
+the source and target domains, which is termed “domain
+shift” problem [7]–[9]. To tackle this problem, various do-
+main adaptation algorithms [10], [11] have been proposed
+to perform knowledge transfer by reducing inter-domain
+distribution discrepancy. To avoid intensive burden of data
+annotation, unsupervised domain adaptation has achieved
+much progress [12]–[15]. As illustrated in Fig. 1 (a), unsuper-
+vised domain adaptation aims to transfer knowledge from a
+labeled source domain to a target domain without accessing
+any target label information.
+Existing deep learning studies on unsupervised domain
+adaptation highly depend on the accessibility of source data,
+which is usually limited in practical scenarios due to the
+following possible reasons. (1) Data privacy protection. Many
+source datasets containing confidential information, such as
+medical and facial data, are not available to third parties
+due to privacy and security protection. (2) Data storage and
+•
+Y. Fang, P.-T. Yap, W. Lin and M. Liu are with the Department
+of Radiology and Biomedical Research Imaging Center, University of
+North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States.
+H. Zhu is with the Department of Biostatistics, University of North
+Carolina at Chapel Hill, NC 27599, USA. Corresponding author: M. Liu
+(mxliu@med.unc.edu).
+transmission cost. The storage and transmission of large-
+scale source datasets, such as ImageNet [16], could bring
+much economic burden. (3) Computation burden. Training on
+extremely large source datasets requires high computational
+resources, which is not practical, especially in real-time
+deployment cases. Thus, there is a high demand for source-
+free unsupervised domain adaptation (SFUDA) methods
+that transfer a pre-trained source model to the unlabeled
+target domain without accessing any source data [17]–[20].
+Many promising SFUDA algorithms have been devel-
+oped recently to address problems in the fields of seman-
+tic segmentation [21], image classification [22], object de-
+tection [23], face anti-spoofing [24], etc. A comprehensive
+review of current studies on SFUDA as well as an outlook
+on future research directions are urgently needed. Liu et
+al. [25] present a review on data-free knowledge transfer,
+where SFUDA only accounts for part of the review and
+the taxonomy of SFUDA is generally rough. And a large
+number of relevant studies have emerged in the past year,
+but the related papers are not included in that survey. In
+addition, their work does not cover commonly used datasets
+in this research field.
+To fill the gap, in this paper, we provide a timely
+and thorough literature review of existing deep learning
+studies on source-free unsupervised domain adaptation.
+Our goal is to cover SFUDA studies of the past few years
+and provide a detailed and systematic SFUDA taxonomy.
+Specifically, we classify existing SFUDA approaches into
+two broad categories: (1) white-box SFUDA as shown in
+Fig. 1 (b) and (2) black-box SFUDA as illustrated in Fig. 1 (c).
+The difference between them lies in whether the model
+parameters of the pre-trained source model are available
+or not. Based on different learning strategies they use, we
+further subdivide white-box and black-box SFUDA methods
+arXiv:2301.00265v1  [cs.CV]  31 Dec 2022
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+2
+(a) Conventional UDA
+(b) White-box SFUDA
+Unlabeled 
+Target 
+Data
+UDA
+Tunable Source Parameters
+Source Model
+(c) Black-box SFUDA
+API
+UDA
+Untunable Source Parameters
+Source Model
+UDA
+Labeled 
+Source 
+Data
+Fig. 1. Illustration of (a) conventional unsupervised domain adaptation (UDA), (b) white-box source-free UDA (SFUDA), and (c) black-box SFUDA.
+Compared with (a) conventional UDA that relies on labeled source data {XS, YS} and unlabeled target data XT , (b, c) SFUDA performs knowledge
+transfer by directly leveraging a pre-trained source model ΦS and unlabeled target data XT . The difference between (b) white-box SFUDA and (c)
+black-box SFUDA lies in whether the learnable parameters of the source model ΦS are accessible or not. API: application programming interface.
+ Black-Box SFUDA 
+Self-Supervised Knowledge Distillation
+Pseudo-Label Denoising
+ White-Box SFUDA 
+Model Fine-Tuning
+Semi-Supervised Knowledge Distillation
+Domain Alignment via Statistics
+Contrastive Learning
+Uncertainty-Guided Adaptation
+Hidden Structure Mining
+Source-Free 
+Unsupervised 
+Domain Adaptation 
+(SFUDA)
+Data Generation
+Domain Image Generation
+Domain Distribution Generation
+Future Outlook
+Multi-Source/Target Domain Adaptation
+Test-Time Domain Adaptation
+Open/Partial/Universal-Set Domain Adaptation
+Flexible Target Model Design
+Cross-Modality Domain Adaptation
+Continual/Lifelong Domain Adaptation
+Semi-Supervised Domain Adaptation
+Generative Distribution Alignment
+Fig. 2. Taxonomy of existing source-free unsupervised domain adaptation (SFUDA) methods, as well as future outlook.
+into finer categories, and the overall taxonomy is shown in
+Fig. 2. Moreover, we discuss the challenges and insight for
+methods in each category, provide a comprehensive compar-
+ison between white-box and black-box SFUDA approaches,
+summarize commonly used datasets in this field as well
+as popular techniques to improve model generalizability
+across different domains. We have to point out that SFUDA
+is still under vigorous development, so we further discuss
+the main challenges and provide insights into potential
+future directions accordingly.
+The rest of this survey is organized as follows. Section 2
+and Section 3 review existing white-box and black-box
+SFUDA methods, respectively. In Section 4, we compare
+white-box and black-box SFUDA and present useful strate-
+gies to improve model generalization. Section 5 discusses
+challenges of existing studies and future research directions.
+Finally, we conclude this paper in Section 6.
+2
+WHITE-BOX
+SOURCE-FREE
+UNSUPERVISED
+DOMAIN ADAPTATION
+Denote ΦS as the source model well-trained based on the
+labeled source domain {XS, YS}, where XS and YS repre-
+sent source data and the corresponding label information,
+respectively. Denote {XT } as the unlabeled target domain
+with only target samples XT . The goal of SFUDA is to learn
+a target model ΦT for improved target inference based on
+the pre-trained source model ΦS and unlabeled target data
+XT . In the setting of white-box source-free domain adaptation,
+the source data (i.e., XS and YS) cannot be accessed but the
+training parameters of the source model ΦS are available. As
+shown in the upper middle of Fig. 2, existing white-box
+SFUDA studies can be divided into two categories: Data
+Generation Method and Model Fine-Tuning Method, with
+details elaborated as follows.
+2.1
+Data Generation Method
+2.1.1
+Domain Image Generation
+Many studies aim to generate source-like image data and
+achieve cross-domain adaptation by readily applying stan-
+dard unsupervised domain adaptation techniques. Based
+on different image generation strategies, these studies can
+be divided into the following three subcategories: (1) batch
+normalization statistics transfer, (2) surrogate source data
+construction, and (3) generative adversarial network (GAN)
+based image generation.
+(1) Batch Normalization Statistics Transfer. Consider-
+ing that batch normalization (BN) stores the running mean
+and variance for a mini-batch of training data in each layer
+of a deep learning model, some studies [26]–[28] explicitly
+leverage such BN statistics for image style transfer, as il-
+lustrated in Fig. 3. For instance, Yang et al. [26] generate
+source-like images via a two-stage coarse-to-fine learning
+strategy. In the coarse image generation step, BN statistics
+stored in the source model are leveraged to preserve the
+style characteristics of source images and also maintain the
+content information of target data. In the fine image genera-
+tion step, an image generator based on Fourier Transform
+is developed to remove ambiguous textural components
+of generated images and further improve image quality.
+With generated source-like images and given target images,
+a contrast distillation module and a compact consistency
+measurement module are designed to perform feature-level
+and output-level adaptation, respectively. Similarly, Hou et
+al. [27] perform style transfer by matching BN statistics of
+generated source-style image features with those saved in
+
+(Xs,Ys]X
+TΦ
+SSOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+3
+Target Data
+Noise
+Source Model
+Source Model
+Source-like Data
+UDA
+Style Transfer via BN matching
+& Content Preservation
+Fig. 3. Illustration of Batch Normalization Statistics Transfer methods
+for source image generation. By matching batch normalization (BN)
+statistics between the upper and lower branches, source-like data can
+be generated by preserving the target content but with source style. Un-
+supervised domain adaptation (UDA) can then be performed between
+source-like data and target data.
+the pre-trained source model for image translation. Hong et
+al. [28] generate source-like images by designing a style-
+compensation transformation architecture guided by BN statis-
+tics stored in the source model and the generated reliable
+target pseudo-labels.
+(2) Surrogate Source Data Construction. To compensate
+for the inaccessible source domain, some studies [29]–[33]
+construct surrogate/proxy source data by selecting appropriate
+samples from the target domain directly, as illustrated in Fig. 4.
+For example, Tian et al. [29] construct pseudo source sam-
+ples directly from the provided target samples under the
+guidance of a designed sample transport rule. The adaptation
+step and sample transport learning step are performed alter-
+nately to refine the approximated source domain and attain
+confident labels for target data, thus achieving effective
+cross-domain knowledge adaptation. Ding et al. [30] build a
+category-balanced surrogate source domain using pseudo-
+labeled target samples based on a prototype similarity mea-
+surement. During model adaptation, intra-domain and inter-
+domain mixup regularizations are introduced to transfer
+label information from the proxy source domain to the target
+domain, as well as simultaneously eliminate negative effects
+caused by noisy labels. Ye et al. [31] select target samples
+with high prediction confidence to construct a virtual source
+set that mimics source distribution. To align the target and
+virtual domains, they develop a weighted adversarial loss
+based on distribution and an uncertainty measurement to
+achieve cross-domain adaptation. Moreover, an uncertainty-
+aware self-training mechanism is proposed to iteratively
+produce the pseudo-labeled target set to further enhance
+adaptation performance. Du et al. [32] construct a surrogate
+source domain by first selecting target samples near the
+source prototypes based on an entropy criterion, and then
+enlarging them by a mixup augmentation strategy [34].
+The adversarial training is then used to explicitly mitigate
+cross-domain distribution gap. Yao et al. [33] simulate proxy
+source domain by freezing the source model and minimiz-
+ing a supervised objective function for optimization. For the
+simulated source set, global fitting is enforced by a model
+gradient based equality constraint, which is optimized by an
+alternating direction method of multipliers algorithm [35].
+(3) GAN-based Image Generation. Instead of approx-
+Target Data
+Surrogate Source 
+Data Construction
+Surrogate Source Data
+UDA
+Fig. 4. Illustration of Surrogate Source Data Construction methods for
+source data generation. These methods first construct surrogate/proxy
+source data by selecting appropriate samples from the target domain
+and then perform standard unsupervised domain adaptation (UDA).
+imating the source domain directly using existing target
+data, Kurmi et al. [36] simulate the source data by training
+a GAN-based generator, as illustrated in Fig. 5. Specifically,
+they first use a parametric conditional GAN to generate la-
+beled proxy source data by treating the source classifier as
+an energy based function. Then, they learn feature patterns
+that are invariant across two domains via standard adver-
+sarial learning for further adaptation. Hou et al. [37] also
+update an image generator framework but they aim to translate
+target images into the source-style ones instead of using the
+latent noise as in [36]. In their method, the knowledge adap-
+tation is achieved by training 1) a knowledge distillation loss
+that mitigates the difference between features of newly gen-
+erated source-style images and those of target images, and
+2) a relation-preserving loss that maintains channel-level
+relationship across different domains. Li et al. [38] propose a
+GAN-embedded generator conditioned on a pre-defined label
+to generate target-style data. By incorporating real target
+samples, the learnable parameters of the generator and the
+adapted model can be updated in a collaborative manner.
+Moreover, two constraints, i.e., weight regularization and
+clustering-based regularization, are utilized during model
+adaptation to preserve source knowledge and ensure high-
+confident target prediction, respectively.
+2.1.2
+Domain Distribution Generation
+Instead of generating source-like images directly, some stud-
+ies propose to align feature prototypes or feature distribu-
+tion of source data [39]–[43] with those in the target domain.
+Specifically, Qiu et al. [39] generate feature prototypes for
+each source category based on a conditional generator and
+produce pseudo-labels for the target data. The cross-domain
+prototype adaptation is achieved by aligning the features
+derived from pseudo-labeled target samples to source pro-
+totype with the same category label via contrastive learn-
+ing. Tian et al. [40] construct a virtual domain by sim-
+ply sampling from an approximated gaussian mixture model
+(GMM) to mimic unseen source domain distribution. In
+terms of adaptation procedure, they reduce the distribution
+gap between the constructed virtual domain and the target
+domain via adversarial training, thus bypassing inaccessible
+source domain. Their practice is based on the assumption
+that the feature prototype of each category can be mined
+from each row of the source classifier’ weights [44]. With
+the same assumption, Ding et al. [41] leverage such source
+classifier weights and reliable target pseudo-labels derived
+by spherical k-means clustering to estimate source feature
+distribution. After that, proxy source data can be sampled
+from the estimated source distribution, and a conventional
+domain adaptation strategy [45] is used to explicitly perform
+
+X
+TΦ
+SX
+TSOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+4
+GAN Generator
+0
+UDA
+Generated 
+Source Data
+Target Data
+Noise
+Pre-defined Label
+Source Model
+Fig. 5.
+Illustration of Generative Adversarial Network (GAN) based
+Image Generation methods for source data generation. Typically, a pre-
+defined label and random noise act as the inputs of a GAN-based gen-
+erator. By utilizing the pre-trained source model, they synthesize source
+data for cross-domain adaptation. LCE: Cross-entropy loss function.
+cross-domain feature distribution alignment. Stan et al. [42],
+[43] propose to first generate a prototypical distribution
+representing the source data in an embedding feature space
+via GMM, and then perform source-free adaptation by
+enforcing distribution alignment between source and target
+domains via sliced Wasserstein distance [46].
+2.1.3
+Challenges and Insight
+We classify existing domain image generation methods for
+SFUDA into three subcategories. We present the challenges
+of methods in each subcatetory and our insights below.
+• Among the above-mentioned three subcategories, the first
+one (i.e., batch normalization statistics transfer) explicitly
+performs BN statistics matching between source and tar-
+get domains for style transfer. Since the BN statistics
+of the source model are off-the-shelf, these methods are
+generally efficient and don’t require complex model train-
+ing. However, BN statistics mainly focus on keeping the
+style features while the content information cannot be
+well preserved. Therefore, this strategy is more applicable
+to scenarios where the contextual structure of images
+between source and target domains does not differ too
+much. It may not show good adaptation performance,
+e.g., from a natural image to a cartoon image, since the
+content information has significant changes. Note that BN
+statistics transfer can also be used as a pre-processing step
+in source-free domain adaptation, and it can be combined
+with other strategies, e.g., circular learning [28], for more
+effective knowledge transfer.
+• Methods in the second subcategory (i.e., surrogate source
+data construction) aim to approximate the proxy source
+domain using appropriate target samples directly, fol-
+lowed by conventional unsupervised domain adaptation.
+Their application is quite broad, including semantic seg-
+mentation [31], object recognition [30], [32], [33], image
+classification [29], and digital recognition [29], [32]. In
+general, methods in this group are straightforward and
+computation-efficient by avoiding introducing extra hy-
+perparameters, which is different from generative models.
+However, because the proxy source samples are directly
+selected from the target domain, these generated source
+data may not effectively represent the original source
+domain. Moreover, how to effectively select informative
+target data for source data approximation is an important
+topic to be investigated. Some studies have proposed var-
+ious strategies for target data selection based on entropy
+measurement [31], source prototype [30], [32], aggregated
+source decision boundary [29], and equality constrained
+optimization [33]. This is still an open but very interesting
+future direction. For multi-source settings, it is promising
+to study which source predictor(s) we should refer to for
+effective target data selection.
+• Methods in the third category (i.e., GAN-based image gen-
+eration) typically synthesize images based on a generative
+model. Since the generator can model underlying complex
+distribution of source data with given random noise,
+GAN-based models generally create more diverse images
+compared with methods in second category (i.e., surro-
+gate source data construction). However, these methods
+introduce additional frameworks and learnable parame-
+ters (e.g., generators and discriminators), which may cost
+more computation resources. By comparing experimental
+results, we find the surrogate source data construction
+methods [32], [33] generally outperform the GAN-based
+generators [36], [38]. The possible reason may be that the
+constructed source data in the former are closer to real
+data distributions, while those recovered in GAN-based
+methods usually suffer from a mode collapse problem [30]
+that leads to low-quality images. Note that the mode col-
+lapse problem can be partly mitigated by using a carefully
+tuned learning rate, manifold-guided training [47], and
+virtual mapping [48], which is worth exploring further.
+Different from image generation methods (Section 2.1.1)
+that directly generate source/target-like images, the dis-
+tribution generation methods (Section 2.1.2) generate fea-
+ture prototype/distribution to achieve cross-domain feature
+alignment. By comparing the reported experimental results,
+we find that the distribution generation approaches [39]–
+[41] usually outperform the GAN-based image generation
+method [38]. And surrogate source data construction meth-
+ods [30], [32] usually show superior performance compared
+with the distribution generation methods [39], [40]. The
+underlying reason could be that the source distributions
+directly derived from the existing target data [30], [32] are
+more accurate and stable than the approximated ones [39],
+[40]. How to drive the approximated source distribution to
+the real one can be further explored in the future.
+2.2
+Model Fine-Tuning Method
+Instead of generating source-like data for standard unsuper-
+vised domain adaptation, many studies attempt to fine-tune
+a pre-trained source model by exploiting unlabeled target
+data in a self-supervised training scheme. Based on differ-
+ent strategies for fine-tuning the source model, we divide
+existing studies into five subcategories: (1) self-supervised
+knowledge distillation, (2) domain alignment via statistics,
+(3) contrastive learning, (4) uncertainty-guided adaptation,
+and (5) hidden structure mining methods, as shown in
+Fig. 2. More details are introduced in the following.
+2.2.1
+Self-Supervised Knowledge Distillation
+Many studies [22], [49]–[55] transfer knowledge learned
+from source data to the target model via knowledge distil-
+lation in a self-supervised manner, as illustrated in Fig. 6. In
+these works, most of them [22], [49]–[52] achieve source-free
+domain adaptation via a mean-teacher scheme for knowledge
+transfer [56], where the target model not only learns from
+unseen target domain but also well preserves source model
+
+X
+TΦ
+SLCESOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+5
+Aug-α
+Aug-β
+Teacher Network
+Student Network
+EMA
+LKD
+Source Model
+Initialize
+Target Data
+Initialize
+Fig. 6. Illustration of Self-Supervised Knowledge Distillation methods
+for source-free unsupervised domain adaptation. With target data from
+different augmentations as inputs, a teacher-student framework is uti-
+lized to exploit target features, where parameters of teacher network are
+usually exponential moving average (EMA) of those of student network.
+Aug-α and Aug-β denote two data augmentation methods (e.g., flip,
+rotation, shift, noise addition, distortion, etc), respectively. LKD: Knowl-
+edge distillation loss function.
+information. For instance, Liu et al. [49] propose a self-
+supervised distillation scheme for automatic polyp detec-
+tion. By means of keeping output consistency of weak and
+strong augmented polyp images, source knowledge is im-
+plicitly transferred to the target model with a mean teacher
+strategy [56]. Besides, a diversification flow paradigm is
+designed to gradually eliminate the style sensitivity among
+different domains, further enhancing model robustness to-
+wards style diversification. Yang et al. [50] also propose
+a self-supervised mean-teacher approach for knowledge
+distillation, with a Transformer module [57] embedded.
+This effectively helps the target model focus on object re-
+gions rather than less informative background in an image,
+thus improving model generalizability. Assuming that both
+source and target images are generated from a domain-
+invariant space by adding noise perturbations on each spe-
+cific domain, Xiong et al. [51] establish a super target domain
+via augmenting perturbations based on the original target
+domain. The super and the original target domains are
+fed into a mean-teacher framework, with three consistency
+regularization terms (w.r.t. image, instance, and class-wise)
+introduced for domain alignment. Chen et al. [22] first divide
+the target data into clean and noisy subsets guided by a
+computation loss and regard them as labeled and unlabeled
+examples, and then utilize the mean teacher technique to
+self-generate pseudo-labels for the unlabeled target data for
+domain adaptation.
+Instead of utilizing the conventional one-teacher one-
+student paradigm, Liu et al. [52] construct a multi-teacher
+multi-student framework, where each teacher/student net-
+work is initialized using a public network pre-trained on
+a single dataset. Here, a graph is constructed to model the
+similarity among samples, and such relationship predicted
+by the teacher networks is used to supervise the student net-
+works via a mean-teacher technique. Rather than leverage
+the mean-teacher paradigm that averages student’s weights,
+Yu et al. [53] propose to distill knowledge from teacher to
+student networks by style and structure regularizations, as
+well as physical prior constraints. Instead of employing a
+teacher-student network as the studies mentioned above,
+Tang et al. [54] achieve data-free adaptation through gradual
+knowledge distillation. Specifically, they first generate pseudo-
+labels via a constructed neighborhood geometry, and then
+Target Data
+Source Model
+Stored Source 
+Statistics
+Target Model
+Derived Target 
+Statistics
+Statistics Discrepancy
+Minimization
+Fig. 7. Illustration of Domain Alignment via Statistics methods for
+source-free unsupervised domain adaptation. The corresponding meth-
+ods leverage batch statistics stored in the pre-trained source model
+to approximate the distribution of inaccessible source data, and then
+perform cross-domain adaptation by reducing distribution discrepancy
+between source and target domains.
+use pseudo-labels obtained from the latest epoch to super-
+vise the current training epoch for knowledge transfer.
+2.2.2
+Domain Alignment via Statistics
+Many studies [58]–[64] leverage batch statistics stored in the
+pre-trained source model to approximate the distribution
+of inaccessible source data, and then perform cross-domain
+adaptation by reducing distribution discrepancy between
+source and target domains, as demonstrated in Fig. 7. For
+example, Ishii et al. [58] approximate feature distribution
+of inaccessible source data by using batch normalization
+statistics (mean and variance) saved in the pre-trained source
+model. Then, Kullback-Leibler (KL) divergence is utilized
+to minimize the distributional discrepancy between source
+and target domains, thus achieving domain-level alignment.
+Inspired by [65], [66], Paul et al. [60] update the mean and
+variance of BatchNorm [67] or InstanceNorm [68] of the
+pre-trained model based on unseen target data. Not limited
+to matching low-order batch-wise statistics (e.g., mean and
+variance), Liu et al. [59] additionally incorporate high-order
+batch-wise statistics, such as scale and shift parameters, to
+explicitly keep cross-domain consistency. Moreover, they
+quantify each channel’s transferability based on its inter-
+domain divergence and assume that the channels with
+lower divergence contribute more to domain adaptation.
+Fan et al. [61] propose to align domain statistics adaptively by
+modulating a learnable blending factor. By minimizing the
+total objective function, each BN layer can dynamically ob-
+tain its own optimal factor, which controls the contribution
+of each domain to BN statistics estimation. The methods
+mentioned above are all based on Gaussian-based statistics
+domain alignment, while Eastwood et al. [62] attempt to
+align histogram-based statistics of the marginal feature dis-
+tributions of the target domain with those stored in the
+pre-trained source model, thus well extending adaptation
+to non-Gaussian distribution scenarios.
+2.2.3
+Contrastive Learning
+Many contrastive learning studies [19], [24], [69]–[72] per-
+form data-free adaptation, which helps the target model
+capture discriminative representations among unlabeled
+target data. The main idea is to pull instances of similar
+categories closer and push instances of different categories away
+in feature space, as illustrated in Fig. 8.
+
+X
+TΦ
+SAug-QAug-βL
+KDΦ
+SX
+TSOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+6
+Before Adaptation
+After Adaptation
+Pull Close
+Push Apart
+Target Data
+Fig. 8. Illustration of Contrastive Learning methods for source-free un-
+supervised domain adaptation. These methods exploit discriminative
+representations among unlabeled target data by pulling instances of
+similar categories closer and pushing instances of different categories
+away in feature space.
+For instance, Xia et al. [69] first adaptively divide tar-
+get instances into source-similar and source-dissimilar sets,
+and then design a class-aware contrastive module for cross-
+set distribution alignment. The idea is to enforce the com-
+pactness of target instances from the same category and
+reduce cross-domain discrepancy, thus prompting effective
+knowledge transfer from the source model to target data.
+Wang et al. [70] present a cross-domain contrastive learning
+paradigm, which aims to minimize the distance between an
+anchor instance from one domain and instances from other
+domains that share the same category as the anchor. Due to
+the unavailability of source data, they utilize source proto-
+typical representations, i.e., weight vectors in the classifier
+layer of a pre-trained source model, for feature alignment
+across two domains. Huang et al. [19] tackle the data-free
+domain adaptation by taking advantage of the historical
+source hypothesis. Specifically, they propose a historical con-
+trastive instance discrimination strategy to learn from target
+samples by contrasting their learned embeddings generated
+by the currently adapted and historical models. And they
+also design a historical contrastive category discrimination
+strategy to weight pseudo-labels of target data to learn
+category-discriminative target representations, by calculat-
+ing the consistency between the current and historical model
+predictions. The two discrimination strategies help exploit
+historical source knowledge, bypassing the dependence on
+source data. Inspired by [73], Agarwal et al. [71] introduce
+a pair-wise contrastive objective function to reduce intra-
+category distance and meanwhile increase inter-category
+distance based on generated target pseudo-labels. They also
+introduce robust source and target models by taking advan-
+tage of the generated adversarial instances, which facilitates
+robust transfer of source knowledge to the target domain.
+2.2.4
+Uncertainty-Guided Adaptation
+Uncertainty can measure how well the target model fits the
+data distribution [74], and many studies [75]–[82] utilize
+such valuable information to guide target predictions in
+source-free adaptation scenarios (see Fig. 9).
+For instance, Fleuret et al. [75] estimate uncertainty based
+on differences between predicted outputs with and without
+Dropout operation [83]. By minimizing such differences, the
+prediction uncertainty on target data is reduced, meanwhile
+the learnable feature abstractor can be more robust to noise
+perturbations. Lee et al. [76] exploit aleatoric uncertainty by
+encouraging intra-domain consistency between target images
+and their augmented ones and enforcing inter-domain feature
+Target Data XT
+Target Model
+Uncertainty Measurement
+Updated
+e.g.,
Monte Carlo Dropout, 
+Entropy, Confidence, Consistency
+Fig. 9. Illustration of Uncertainty-Guided Adaptation methods for source-
+free unsupervised domain adaptation. These studies utilize uncertainty
+to guide target predictions, and such valuable information can be mea-
+sured by Monte Carlo Dropout, entropy, etc.
+distribution consistency. Chen et al. [77] introduce a predic-
+tion denoising approach for a cross-domain segmentation
+task. In this study, a key component is introducing pixel-
+wise denoising via uncertainty evaluation using Monte Carlo
+Dropout [84], [85], which calculates the standard deviation
+of several stochastic outputs and keeps it under a manually-
+designed threshold. In this way, the noisy pseudo-labels
+can be filtered out, helping improve pseudo-label quality
+to achieve effective adaptation. Xu et al. [78] also propose an
+uncertainty-guided pseudo-labeling denoising scheme, but
+they use soft label correction instead of manually discarding
+unreliable data points. Specifically, they first identify misla-
+beled data points by utilizing a joint distribution matrix [86],
+[87], and then assign larger confident weights to those with
+higher certainty based on Monte Carlo Dropout. Combining
+target data and the corresponding rectified pseudo-labeling,
+a commonly used cross-entropy objective function can be
+leveraged for training the target model. Sharing the similar
+idea, Hegde et al. [79] allocate lower weights for uncer-
+tain pseudo-labels, where the uncertainty is measured by
+prediction variance based on Monte Carlo Dropout [84],
+[85]. Considering that using Monte Carlo Dropout [84]
+for uncertainty estimation requires manual hyperparameter
+adjustment [88], Roy et al. [80] quantify source model’s un-
+certainty using a Laplace approximation [89], [90]. For model
+training, they assign smaller weights to those target samples
+that are farther away from source hypothesis (measured by
+uncertainty), avoiding misalignment of dissimilar samples.
+Pei et al. [81] tackle the uncertainty issue from the perspec-
+tive of improving source model transferability. Specifically,
+they estimate channel-aware transferability of the source
+model to target data based on an uncertainty distance,
+which measures the closeness between target instances and
+source distribution. With the aim of dynamically exploiting
+the source model and target data, the target model obtains
+the source knowledge from the transferable channels and
+neglects those less-transferable ones. Unlike previous stud-
+ies, Li et al. [82] quantify uncertainty using self-entropy and
+propose a self-entropy descent mechanism to seek the optimal
+confidence threshold for robust pseudo-labeling of target
+data. They also leverage false negative mining and mosaic
+augmentation [91] to further eliminate the negative influ-
+ence of noisy labels to enhance adaptation performance.
+2.2.5
+Hidden Structure Mining
+Many studies [20], [92]–[98] take into consideration intrinsic
+feature structures of target domain and update the target
+model via clustering-aware pseudo-labeling. In Fig. 10, we
+illustrate the main idea of hidden structure mining methods.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+7
+Before Adaptation
+Class Centroid
+Iteratively Update 
+Clustering Centroid  
+After Adaptation
+Target Data
+Fig. 10. Illustration of Hidden Structure Mining methods for source-free
+unsupervised domain adaptation. These methods take into consider-
+ation intrinsic feature structures of target domain and iterate between
+target model refinement and clustering centroid update.
+For example, Yang et al. [20] observe that target data can
+intrinsically form a certain cluster structure that can be used
+for domain adaptation. Specifically, they estimate affinity
+among target data by taking into account the neighborhood
+patterns captured from local-, reciprocal-, and expanded-
+neighbors. Source-free adaptation is achieved by encour-
+aging consistent predictions for those with high affinity.
+Similarly, Yang et al. [92] also exploit neighborhood struc-
+ture information of target data. They propose a local struc-
+ture clustering strategy to encourage prediction consistency
+among k-nearest target features, thus pushing target data
+with semantically similar neighbors closer. Tang et al. [93]
+leverage semantic constraints hidden in geometric structure
+among target data to encourage robust clustering based
+on a cognition mechanism [99]. Source hypothesis transfer
+(SHOT) [94] and SHOT++ [95] attempt to mine the feature
+structure of the target domain, but they cannot fully ex-
+ploit the meaningful context since the used self-supervised
+pseudo-labeling does not take into account each dimen-
+sion’s covariance in the feature space. To address this issue,
+Lee et al. [96] utilize GMM in the target domain to obtain
+data structure, and design a joint model-data structure score to
+concurrently exploit source and target knowledge. Yang et
+al. [97] propose a novel neighborhood structure cluster-
+ing method, which encourages intra-cluster target features
+closer and meanwhile disperses those inter-cluster target
+predictions far away. Li et al. [98] utilize neighbor structure
+information from a new aspect by proposing a generic
+and model smoothness-assisted Jacobian norm regularization
+term, which is used to manipulate the consistency between
+each target instance and its neighbors. This Jacobian norm
+regularizer can be easily plugged into existing source-free
+domain adaptation frameworks for boosting performance.
+Different from the above mentioned methods, some
+studies tackle source-free domain adaptation from other
+perspectives. Li et al. [100] achieve data-free adaptation from
+an adversarial-attack aspect, which aims to generate adversar-
+ial target instances by adding diverse perturbations to attack
+the target model. Then, mutual information maximization is
+performed between representations extracted by the source
+and target model for the same target instance. The above
+two steps are performed alternatively, by which the domain-
+invariant source knowledge can be preserved and the rich
+target patterns can be well explored. Instead of explor-
+ing domain-invariant features for cross-domain knowledge
+transfer, Wang et al. [101] mine domain-invariant parameters
+stored in the source model. They assume that only partial
+domain-invariant parameters of the source model contribute
+to domain adaptation, and their goal is to capture such pa-
+rameters while penalizing the domain-specific ones. Liang et
+al. [102] explore source-free adaptation from the perspec-
+tive of minimum centroid shift, with the aim of searching a
+subspace where target prototypes are mildly shifted from
+source prototypes. An alternating optimization scheme is
+leveraged for model convergence and target pseudo-label
+update. Inspired by maximum classifier discrepancy [14],
+Yang et al. [103] introduce an auxiliary bait classifier for cross-
+domain feature alignment combined with the source anchor
+classifier. These two classifiers aim to collaboratively push
+uncertain target representations to the correct side of the
+source classifier boundary.
+2.2.6
+Challenges and Insight
+We classify existing model fine-tuning methods for SFUDA
+into five subcategories. The challenges of methods in each
+subcatetory and our insights are presented below.
+• The methods in the first subcategory, i.e., self-supervised
+knowledge distillation, interpret source-free domain adap-
+tation as a knowledge extraction and transfer process,
+aiming to learn domain-invariant feature representations.
+Most exiting studies transfer source knowledge to the
+target model via a mean teacher strategy [56], where
+teacher weights are an exponential moving average of
+student weights. Hence, model parameters of both teacher
+and student networks are tightly coupled, which may
+lead to a performance bottleneck. A possible solution is to
+introduce a dual-student framework and let one student
+learn features flexibly, which may disentangle teacher-
+student weights to some extent [104].
+• The second subcategory, i.e., domain alignment via statistics,
+leverages batch statistics stored in a pre-trained source
+model to approximate distribution of inaccessible source
+data. Compared with other categories, these statistics-
+based methods are lightweight and prone to generalize to
+other tasks, since they require only a few update steps of
+batch-wise statistics parameters and are potentially appli-
+cable to real-time deployment [64]. However, they are not
+suitable for problems that use deep network architectures
+without batch normalization layers.
+• The methods in the third subcategory, i.e., contrastive learn-
+ing, aim to bring similar-class samples closer and push
+dissimilar-class samples apart based on generated tar-
+get pseudo-labels. Therefore, if the pseudo-labels contain
+much noise, these methods may suffer from substantial
+performance degradation. Moreover, a memory bank is
+usually required to store the similarity relationship be-
+tween current and historical feature representations of
+target data, which could bring memory burden. It is
+interesting to investigate the storage- and transmission-
+efficient contrastive learning strategies in source-free set-
+tings. In addition, several recent studies [105], [106]
+have shown that data pair construction is crucial for
+effective contrastive learning. One solution is utilizing
+contrastive information between target data and their
+augmented ones. Previous studies [107] often use either
+strong or weak transformations for data augmentation,
+where strong augmentations mostly distort the structures
+of original images (e.g., shape distortion) while weak
+augmentations usually limit transformations to preserve
+the images’ structures (e.g., flip). Here we propose to
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+8
+dynamically mix strong and weak augmentation of target
+data, which may help learn more robust representations.
+• The methods in fourth subcategory, i.e., uncertainty-guided
+adaptation, focus on reducing prediction uncertainty of tar-
+get data. Many studies [77], [78] use Monte Carlo Dropout
+for uncertainty estimation, but this technique requires
+specialized network architecture design and model train-
+ing, bringing troublesome hyperparameter tuning [88]. A
+recent study [77] points out that their method can only
+handle problems with minor domain shift, and performs
+poorly on problems with severe domain shift. It is inter-
+esting to explore this challenging problem in the future.
+• The last subcategory, i.e., hidden structure mining, considers
+intrinsic clustering structure of the target domain, assum-
+ing that geometric structure of target data may provide
+informative context [93]. The advantage of these methods
+is that no auxiliary frameworks are required, and thus,
+they can be easily incorporated into other adaptation
+frameworks. However, these methods have at least three
+disadvantages. (1) Most existing studies need to iterate
+between feature clustering and model update, which may
+hinder training efficiency and cause a memory burden.
+(2) These methods may be infeasible to handle extremely
+large-scale datasets due to the difficulty of saving global
+latent feature embeddings of the whole dataset [108].
+(3) Most studies construct target geometric structures in
+Euclidean space, which may not be suitable for problems
+with non-Euclidean data such as graphs. Thus, how to
+improve training efficiency and deal with the large-size
+dataset, as well as mining geometry information of non-
+Euclidean data deserve further research.
+From the application perspective, computation-efficient
+approaches are more applicable for pixel-wise semantic
+segmentation tasks, which require higher resources com-
+pared with classification tasks. And those memory-intensive
+approaches such as contrastive learning may be not suitable
+for semantic segmentations. Moreover, it is worth noting
+that data generation methods detailed in Section 2.1 can
+be used in conjunction with the model fine-tuning methods
+described in this section. For instance, one can first generate
+a virtual source domain by selecting appropriate target
+samples, and thus a standard unsupervised domain adap-
+tation framework could be applied. To further exploit target
+information, we then take account of geometric structure of
+target samples and generate corresponding target pseudo-
+labels to fine-tune the target model. These two steps can be
+optimized iteratively, helping generate more representative
+source domain and refine the target model.
+3
+BLACK-BOX
+SOURCE-FREE
+UNSUPERVISED
+DOMAIN ADAPTATION
+Different from white-box methods, in the setting of black-box
+source-free domain adaptation, both the source data {XS, YS}
+and detailed parameters of the source model ΦS are not accessi-
+ble. Only the hard or soft model predictions of the target
+data XT from the source model ΦS are leveraged for do-
+main adaptation. Depending on the utilization of the black-
+box predictor, the existing black-box SFUDA studies can
+be mainly divided into three categories: Self-Supervised
+Knowledge Distillation, Pseudo-Label Denoising, and
+Generative Distribution Alignment methods, with details
+introduced below.
+3.1
+Self-Supervised Knowledge Distillation
+Some studies [109]–[114] construct a teacher-student-style
+network architecture with knowledge distillation to trans-
+late the source knowledge to the target domain in a self-
+supervised manner. For instance, Liang et al. [109], [110]
+enforce output consistency between a source model (i.e.,
+teacher) and a customized target model (i.e., student) via
+a self-distillation loss. Specifically, a memory bank is first
+constructed to store the prediction of each target sample
+based on the black-box source model. This source model
+then acts as a teacher to maintain an exponential moving
+averaging of source and target prediction following [115],
+[116]. Additionally, structural regularization on the target
+domain is further incorporated during adaptation for more
+effective knowledge distillation. Similarly, Liu et al. [111],
+[112] employ an exponential mixup decay scheme to explicitly
+keep prediction consistency of source and target domains,
+thus gradually capturing target-specific feature representa-
+tions and obtaining the target pseudo-labels. Xu et al. [113]
+extend the teacher-student paradigm from image analysis
+to more challenging video analysis, where not only spa-
+tial features but also temporal information are taken into
+consideration during domain adaptation. For knowledge
+distillation, the target model is regarded as a student, which
+aims to learn similar predictions generated by a teacher (i.e.,
+source) model. The teacher model is meanwhile updated
+to maintain an exponential moving averaging prediction.
+Instead of distilling knowledge between source and target
+domains, Peng et al. [114] transfer knowledge between the
+target network and its subnetwork in a mutual way, where
+the subnetwork is a slimmer version generated from the
+original target network following Yang et al. [117]. And
+target features are extracted by leveraging multi-resolution
+input images, helping improve the generalization ability of
+the target network. Moreover, a novel data augmentation
+strategy, called frequency MixUp, is proposed to empha-
+size task-related regions-of-interests while simultaneously
+reducing background interference.
+3.2
+Pseudo-Label Denoising
+Some studies [118], [119] tackle domain shift by carefully
+denoising unreliable target pseudo-labels. For example,
+Zhang et al. [118] combat noisy pseudo-labels via noise rate
+estimation, which first preserves more training samples at
+the start of the training process following [120] and then
+gradually filters out the noisy ones based on their loss
+values as training proceeds. The pseudo-labels are itera-
+tively refined according to a category-dependent sampling
+strategy, encouraging the model to capture more diverse
+representations to improve model generalization ability. Dif-
+ferent from Zhang et al. [118] that only select part of reliable
+target data during model training, Luo et al. [119] take
+into account all target data and rectify noisy pseudo-labels
+from a negative learning aspect. Specifically, their approach
+assigns complementary ground-truth labels for each target
+sample, helping alleviate error accumulation for noisy pre-
+diction. Moreover, a maximum squares objective function is
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+9
+utilized as confidence regularization to prevent the target
+model from being trapped in easy sample training. Yang et
+al. [121] incorporate pseudo-label denoising and self-supervised
+knowledge distillation into a unified framework. Specifically,
+domain knowledge is first distilled from the trained source
+predictor to warm up the target model by an exponential
+moving averaging scheme. The unlabeled target domain
+is then split into two subsets (i.e., easy and hard groups)
+according to their adaptation difficulty [122], and the Mix-
+Match strategy [123] is leveraged to progressively exploit all
+target representations. In this way, the noise accumulation
+is further suppressed, thereby improving the efficacy of
+pseudo-label denoising.
+3.3
+Generative Distribution Alignment
+Different from the above methods, some studies perform
+distribution alignment across domains in a generative way.
+For instance, Yeh et al. [124] perform domain adaptation
+by maximizing the lower bound in variational inference.
+Specifically, they construct a generation path as well as an
+inference path, where the generation path produces a prior
+feature distribution derived from predicted category labels,
+and the inference path approximates a posterior feature
+distribution based on each target instance. The latent dis-
+tribution alignment can be achieved by maximizing the evi-
+dence lower vound in variational inference for cross-domain
+adaptation. Similarly, Yang et al. [125] also construct the
+generation and inference paths, but they achieve adaptation
+via minimizing the upper bound of the prediction error of
+target data in variational inference. Zhang et al. [126] achieve
+source-free adaptation by first building multiple source
+models and then generating a virtual intermediate surrogate
+domain to select target samples with minimum inconsistency
+predicted by the source models. The knowledge transfer
+is achieved by feature distribution alignment between the
+virtual surrogate domain and the target domain based on a
+joint probability maximum mean discrepancy [127].
+3.4
+Challenges and Insight
+In this section, we classify existing black-box SFUDA meth-
+ods into three categories based on how they utilize the noisy
+target predictions. The challenges of each category and our
+insights are presented below.
+• The first category, i.e., self-supervised knowledge distillation,
+aims to gradually transfer source knowledge to a cus-
+tomized target model by enforcing output consistency
+between a teacher (source) and a student (target) network.
+This learning strategy has also been used in white-box
+SFUDA (see Section 2.2.1). The difference is that model
+weights of student networks are accessible in white-box
+SFUDA methods, but not in black-box SFUDA. In black-
+box SFUDA, instead of leveraging any parameter details,
+the teacher network is only updated by source predic-
+tions and historical target predictions. The two items are
+typically weighted by a momentum factor, which helps
+dynamically adjust their contributions. Self-supervised
+knowledge distillation has shown promising performance
+in object recognition [109], semantic segmentation [111],
+and video action recognition tasks [113].
+• The methods in the second category, i.e., pseudo-Label
+denoising, tackle black-box SFUDA from the perspective of
+noisy label rectification. It has shown that a pseudo-label
+denoising approach [118] has inferior performance than
+the self-supervised knowledge distillation method [109].
+The reason may be that the former [118] only focuses on
+noisy prediction itself while neglecting target data struc-
+ture that is well considered in the latter [109]. Considering
+that pseudo-label denoising methods can tackle unbal-
+anced label noise via noise rate estimation, combining
+pseudo-label denoising with self-supervised knowledge
+distillation strategies will be a promising future direc-
+tion, especially in class-imbalance scenarios. Moreover,
+if the black-box predictor only provides one-hot hard
+predictions instead of probability predictions, the utility
+of methods in this subcategory will be greatly reduced.
+The reason is that the noise rate cannot be well estimated
+in practice, e.g., there is nearly no difference between the
+output of [0.45, 0.55] and that of [0.05, 0.95] because the
+source predictor produces the same output (i.e., [0, 1]).
+• The third category, i.e., generative distribution alignment, at-
+tempts to perform domain adaptation by minimizing fea-
+ture distribution discrepancy across the source and target
+domains. Since source distribution is inaccessible in black-
+box models, some generative approaches are utilized to
+generate such reference distribution for target data to
+align with, including variational autoencoder [124], [125]
+and surrogate source domain construction [126]. These
+methods are more suitbale for recognition/classification
+tasks, but less suitable for semantic segmentation tasks.
+For example, generating surrogate feature distribution of
+an object (e.g., car) is usually easier than that of a semantic
+scene (e.g., cityscape), since the latter contains different
+objects and thus the pixel-wise neighborhood relationship
+is difficult to model in practice.
+Besides the strategies proposed above, it is also crucial
+to build a general and robust black-box source model, with
+which the target predictions tend to be more accurate.
+To achieve that, one possible solution is augmenting the
+diversity of source data (e.g., adding some perturbation)
+before constructing the source model, which may eliminate
+style discrepancy between two domains, thus improving the
+generalization ability of the source model. Another solution
+is using soft probability labels instead of hard one-hot
+labels (e.g., [0, 1]) for model training, which prevents the
+source model from being over-confident and helps enhance
+its generalizability. Compared to white-box methods, there
+are relatively few black-box SFUDA methods as well as
+benchmark datasets, which needs to be further explored.
+4
+DISCUSSION
+In this section, we first compare the white-box and black-
+box SFUDA methods and then summarize several useful
+strategies to improve model generalizability. We also list
+datasets commonly used in the field in Table 1.
+4.1
+Comparison of White-Box and Black-Box SFUDA
+By comparing existing white-box and black-box SFUDA
+methods, we have the following interesting observations.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+10
+TABLE 1
+Commonly used datasets for evaluating the performance of source-free unsupervised domain adaptation (SFUDA) approaches.
+Dataset
+Domain #
+Instance #
+Category #
+Description
+Digit Recognition
+Digits-Five [128]
+5
+215,695
+10
+MNIST [129], SVHN [130], USPS [131], MNIST-M [13], Synthetic Digits [13]
+Semantic Segmentation
+Segmentation datasets
+4
+45,766
+-
+GTA5 [132], Cityscapes [133], SYNTHIA [134], NTHU [135]
+Object Recognition
+Office-31 [136]
+3
+4,652
+31
+Amazon, Webcam, DSLR
+Office-Home [137]
+4
+15,500
+65
+Artistic, Clip Art, Product, Real-World
+VisDA [138]
+2
+280,000
+12
+Synthetic and real images
+Office-Caltech-10 [139]
+4
+2,533
+10
+Amazon, DSLR, Webcam, Caltech10
+ImageCLEF-DA [140]
+4
+2,400
+12
+Caltech-256 [141], ImageNet ILSVRC2012 [16], PASCAL VOC2012 [142], Bing [143]
+PACS [7]
+4
+9,991
+7
+Art painting, Cartoon, Photo, Sketch
+DomainNet [144]
+6
+600,000
+345
+Clipart, Infograph, Painting, Quickdraw, Real, Sketch
+MiniDomainNet [145]
+4
+140,000
+126
+Clipart, Painting, Real, Sketch
+PointDA-10 [146]
+3
+33,067
+10
+ModelNet [147], ShapeNet [148], ScanNet [149]
+Face Anti-Spoofing
+Face datasets
+4
+7,130
+-
+Replay-Attack [150], OULU-NPU [151], CASIA-MFSD [152], MSU-MFSD [153]
+LiDAR Detection
+LiDAR datasets
+3
+158,510
+-
+Waymo [154], KITTI [155], nuScenes [156]
+Video Action Recognition
+UCF-HMDBfull [157]
+2
+3,209
+12
+UCF101 [158], HMDB51 [159]
+Sports-DA [160]
+3
+40,718
+23
+UCF10 [158], Sports-1M [161], Kinetics [162]
+Daily-DA [160]
+4
+18,949
+8
+ARID [163], HMDB51 [159], Moments-in-Time [164], Kinetics [162]
+Traffic Sign Recognition
+Sign datasets
+2
+151,839
+43
+Syn.Signs [165], GTSRB [166]
+Image Classification
+VLCS [167]
+4
+10,729
+5
+Caltech101 [168], LabelMe [169], SUN09 [170], VOC2007 [142]
+Medical Data
+BraTS2018 [171]
+2
+285
+-
+Cross-disease (high and low grade glioma), cross-modality (T1, T1ce, T2, FLAIR)
+MMWHS [172]
+2
+40
+-
+Cross-modality (magnetic resonance imaging, computed tomography)
+Brain skull stripping [173]
+3
+35
+-
+NFBS [174], ADNI [175], dHCP [176]
+Polyp segmentation
+4
+2,718
+-
+CVC-ClincDB [177], Abnormal Symptoms [178], ETIS-Larib [179], EndoScene [180]
+EEG MI Classification [126]
+4
+528
+2/4
+MI2-2 [181], MI2-4 [181], MI2015 [182], AlexMI [183]
+Prostate segmentation
+2
+682
+-
+NCI-ISBI 2013 Challenge [184], PROMISE12 challenge [185]
+Optic disc&cup segmentation
+3
+660
+-
+REFUGE [186], RIMONE-r3 [187], Drishti-GS [188]
+Autism diagnosis
+4
+411
+2
+NYU, USM, UCLA, and UM of ABIDE dataset [189]
+• Compared with black-box SFUDA that cannot access any
+source parameters, white-box SFUDA is capable of mining
+more source knowledge (e.g., batch statistics) that facili-
+tates more effective domain adaptation.
+• White-box SFUDA methods may suffer from data pri-
+vacy leakage problems [118]. For instance, Yin et al. [190]
+reveal that raw data can be recovered based on source
+image distribution via a deep inversion technique. Using
+a membership inference attack strategy [191], [192], it is
+possible to infer whether a given sample exists or not
+in training dataset, thereby revealing private information.
+The black-box SFUDA can help protect data privacy be-
+cause only application programming interface (API) is
+accessible while detailed model weights are withheld,
+but it may suffer from performance degradation of cross-
+domain adaptation.
+• Most white-box SFUDA methods assume that model ar-
+chitecture is shared between source and target domains,
+while the black-box SFUDA methods try to design task-
+specific target models for knowledge transfer. Such flex-
+ible model design in black-box SFUDA methods is very
+useful for target users with low computation resources,
+since they can design more efficient and lightweight target
+models for domain adaptation.
+• Black-box SFUDA methods neither require data synthe-
+sis nor model fine-tuning, which helps to accelerate the
+convergence process of model training. In contrast, white-
+box methods are usually computationally intensive and
+time-consuming. For instance, it is reported that the com-
+putational cost of a black-box SFUDA method [126] is
+0.83s while that of two competing white-box methods
+are 3.17s [94] and 22.43s [193], respectively, reflecting the
+computation efficiency of black-box SFUDA.
+In summary, when using white-box and black-box
+SFUDA methods, we have to make a trade-off between
+obtaining better performance, protecting confidential infor-
+mation, and reducing computational and memory costs.
+4.2
+Useful Strategies for Improved Generalizability
+To facilitate research practice in this field, we summarize
+several useful techniques that could be used to improve the
+generalizability of learning models for source-free unsuper-
+vised domain adaptation.
+4.2.1
+Entropy Minimization Loss
+Most SFUDA methods utilize an entropy minimization
+loss [194] to reduce uncertainty of model predictions [27],
+[59], [75], [94], [111], [112], [195]–[200]. This simple yet
+effective strategy encourages the model to generate one-hot
+predictions for more confident learning.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+11
+4.2.2
+Diversity Enforcing Loss
+To prevent predicted labels from collapsing to categories
+with larger number of samples, many studies leverage a
+diversity enforcing loss to encourage diverse predictions
+over target domain [80], [94], [196], [201]–[205] . The usual
+practice is to maximize the entropy of empirical label distri-
+bution over the batch-wise average of model predictions.
+4.2.3
+Label Smoothing Technique
+In source-free adaptation studies, a pre-trained source
+model is generally obtained via training on labeled source
+data before adaptation stages. Currently, many studies use
+a label smoothing technique [206], [207] to produce a robust
+source model [20], [30], [94], [101], [208], [209]. This tech-
+nique aims to transform original training labels from hard
+labels (e.g., 1) to soft labels (e.g., 0.95), which prevents the
+source model from being over-confident, helping enhance
+its generalization ability. Also, the experiments have shown
+that label smoothing can encourage closer representations
+of training samples from the same category [206]. With a
+more general and robust source model, it is likely to boost
+adaptation performance on target domain.
+4.2.4
+Model Regularization
+Many regularization terms are utilized in existing SFUDA
+methods by incorporating some prior knowledge. For in-
+stance, an early learning regularization [39], [120], [210] is
+used to prevent the model from over-fitting to label noise. A
+stability regularization [38], [211]–[213] is leveraged to pre-
+vent parameters of the target model to deviate from those
+of the source model. A local smoothness regularization [38],
+[214] is used to encourage output consistency between the
+target model and its noise-perturbed counterpart, helping
+improve robustness of the target model. A mixup regular-
+ization [30], [109], [114], [215], [216] is used to enforce pre-
+diction consistency between original and augmented data,
+which can mitigate the negative influence of noisy labels.
+4.2.5
+Confidence Thresholding
+Many studies leverage pseudo-labeling to train the tar-
+get model in a self-supervised way. Instead of utilizing a
+manually-designed threshold to identify reliable/confident
+pseudo-labels, a commonly used strategy is automatically
+learning the confidence threshold for reliable pseudo-label
+selection [217]. To further tackle the class-imbalance prob-
+lem, some studies [75], [78], [212], [218], [219] propose to
+learn dynamic threshold for each category, which provides
+a fair chance for categories with limited samples to generate
+pseudo-labels for self-training.
+5
+FUTURE OUTLOOK
+5.1
+Multi-Source/Target Domain Adaptation
+To utilize diverse and rich information of multiple domains,
+a few studies [29], [193], [204], [220], [221] propose multi-
+source data-free adaptation to transfer source knowledge
+to the target domain. Tian et al. [29] introduce a sample
+transport learning method, but the proposed model is shal-
+low, and thus cannot handle highly nonlinear feature ex-
+traction. To tackle this problem, several deep learning based
+models [193], [204] are proposed. But they ignore the fact
+that the generated target pseudo-labels may be noisy, which
+may cause training bias when matching target domains with
+large domain gaps. The key to solving problems with multi-
+source domains is quantifying the transferability of different
+source models and utilizing their complementary informa-
+tion for promoting cross-domain adaptation. Even several
+strategies are proposed (e.g., aggregation weight [193] and
+source-specific transferable perception [204]), more explo-
+rations are encouraged to address the problem of negative
+transfer during cross-domain knowledge transfer.
+A few studies [222]–[226] incorporate federated learning
+into domain adaptation scenarios. Federated learning [227]–
+[229] is a decentralized scheme to facilitate collaborative
+learning among multiple distributed clients without shar-
+ing training data or model parameters. The constraint that
+prevents data and parameter transmission across different
+source domains is not required in multi-source-free domain
+adaptation. For instance, federated adversarial domain adapta-
+tion (FADA) introduced by Peng et al. [222] is among the
+first attempts to propose the concept of federated domain
+adaptation, which employs a dynamic attention mecha-
+nism to transfer knowledge from multi-source domains to
+an unlabeled target domain. In this method, each source
+model needs to synchronize with the target domain after
+each training batch, resulting in huge computation costs
+and potential risk of privacy leakage [230]. To tackle this
+problem, Feng et al. [223] introduces a consensus focus
+schema that greatly improves communication efficiency for
+decentralized domain adaptation. Moreover, Song et al. [225]
+utilize a homomorphic encryption approach for privacy pro-
+tection, and Qin et al. [226] introduce a flexible uncertainty-
+aware strategy for reliable source selection. However, cur-
+rent federated learning studies usually produce a common
+model for all clients without considering heterogeneity of
+data distribution of different clients. Therefore, the common
+model cannot adapt to each client adaptively, which may
+affect adaptation performance. It would be very interesting
+to investigate personalized federated learning [231], with
+which current or new clients can easily adapt to their
+own local dataset by performing a few optimization steps.
+Besides, all the methods mentioned above require labeled
+data from multiple sources to train a federated model, in-
+evitably increasing annotation costs. Therefore, approaches
+that effectively exploit unlabeled data from multiple source
+domains in a decentralized way are urgently needed.
+On the other hand, Yao et al. [232] and Shenaj et al. [233]
+have proposed several federated multi-target domain adap-
+tation strategies for transferring knowledge of a labeled
+source server to multiple unlabeled target clients. More ad-
+vanced techniques for federated multi-target domain adap-
+tation are highly desirable, by considering computation
+and communication cost, annotation burden, and privacy
+protection of different target domains.
+5.2
+Test-Time Domain Adaptation
+Most SFUDA approaches require pre-collected unlabeled
+target data for model training, termed “training-time adap-
+tation”. Test-time adaptation [198], [234]–[237] has been
+investigated by adapting the source model to the target
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+12
+domain during inference procedure only. The advantages of
+test-time adaptation are mainly twofold: (1) The adaptation
+process does not need iterative training, which greatly im-
+proves computational efficiency, so the model can be easily
+deployed in an online manner. (2) Without relying on target
+training data, test-time adaptation is expected to be well
+generalized to diverse target domains. Even current studies
+have made promising achievements, there are still some
+problems worth exploring, listed as follows.
+Some studies [198], [238], [239] need to access batch-
+sized (>1) target samples during inference, which can-
+not handle scenarios where target samples arrive one-by-
+one sequentially. Two studies [240], [241] perform image-
+wise adaptation rather than batch-wise adaptation, but they
+cannot deal with cases with large distribution shift. It is
+interesting to explore how to handle the scenarios where
+test instances come from continuous changeable domains
+in the future. Additionally, the solutions on how to adap-
+tively exploit test data can be further explored [242], such
+as adjusting model weights dynamically based on sample
+discrepancy across domains.
+5.3
+Open/Partial/Universal-Set Domain Adaptation
+This survey focuses on close-set source-free domain adapta-
+tion, where the label space of source and target domains is
+consistent. But practical scenarios are much more compli-
+cated when the category shift issue occurs across different
+domains. There are three non-close-set scenarios: (1) open-set
+(Cs ⊂ Ct) problems, (2) partial-set (Cs ⊃ Ct) problems, and
+(3) universal-set (Cs\Ct ̸= ∅ and Ct\Cs ̸= ∅, Cs ⊂ Ct,
+Cs ⊃ Ct) problems, where Cs and Ct denote the category
+label set for source and target domains, respectively.
+Currently, only a few studies [18], [243], [244] attempt
+to handle the category shift problem in source-free adapta-
+tion scenarios, including out-of-distribution data construc-
+tion [18], [243], neighborhood clustering learning [245],
+uncertainty-based progressive learning [246], and mutual
+information maximization [203]. The main idea behind these
+studies is to recognize out-of-source distribution samples
+and improve generalization ability of the source model.
+However, the model performance of existing studies is
+not quite satisfactory due to the inaccessibility of valuable
+category-gap knowledge. One possible solution is to adap-
+tively learn a threshold instead of using a fixed one to
+determine the acceptance/rejection of each target sample
+as a “known” category via some similarity measurement.
+Moreover, some strategies used in non-source-free domain
+adaptation can also be borrowed, such as distribution
+weighted combining rule [247], category-invariant represen-
+tation learning [248], one-vs-all learning scheme [249], and
+global-local optimization [250].
+5.4
+Flexible Target Model Design
+For black-box SFUDA methods, due to unavailability of
+structure and parameters of target model, one usually has
+to manually design a target model. For instance, Liu et
+al. [111] choose a U-Net based framework as the target
+model for segmentation. However, such manually designed
+architectures may be not suitable when adapting to the tar-
+get domain. It is expected that the automatic design of target
+models, e.g., using neural architecture search (NAS) [251]–
+[253], helps improve the learning performance. Consider-
+ing that NAS has recently become a popular strategy for
+searching proper network architectures in deep learning, we
+can integrate it into SFUDA scenarios to find more proper
+and efficient target models. And how to balance the search
+space and search cost of network parameters can be further
+investigated. Moreover, hyperparameters used in NAS (e.g.,
+optimizer strategy, weight decay regularization) should be
+carefully considered since they also have a significant im-
+pact on network performance [254].
+5.5
+Cross-Modality Domain Adaptation
+Existing studies mainly focus on one single modality for
+domain adaptation, while a few studies perform cross-
+modality adaptation in source-free settings [28], [255]. For
+instance, for medical data analysis, the acquisition expense
+of computed tomography (CT) scans is generally less than
+that of magnetic resonance imaging (MRI) scans, hence it
+may greatly reduce annotation cost for a segmentation task
+when transferring a source model trained on CT images to
+MRI scans [28]. Moreover, in computer vision field, it would
+be promising to investigate cross-modality adaptation in
+the future, e.g., image→video, which aims to achieve video
+recognition based on the source model trained on the image
+dataset. Also, how to effectively integrate multi-modality
+(e.g., image, sound, text, and video) data for domain adapta-
+tion in a source-free way is an interesting but not yet widely
+studied problem.
+5.6
+Continual/Lifelong Domain Adaptation
+Most current studies focus on improving adaptation per-
+formance on the target domain while neglecting the perfor-
+mance on source domain, running the risk of catastrophic
+forgetting problems [256]. To address this issue, several
+solutions have been developed from different aspects, such
+as domain expansion [257], historical contrastive learn-
+ing [19], domain attention regularization [92], and model
+perturbation [258], while there is still massive room for
+performance improvement. Inspired by continual/lifelong
+learning [259]–[262], continual domain adaptation has re-
+cently made great progress by investigating gradient reg-
+ularization [263], iterative neuron restoration [264], buffer
+sample mixture [265], etc. The continual domain adaptation
+in source-free settings for mitigation of catastrophic forget-
+ting remains an underdeveloped topic that can be further
+explored in the future.
+5.7
+Semi-Supervised Domain Adaptation
+Source-free domain adaptation in semi-supervised settings
+(i.e., with a few labeled target data involved for model train-
+ing) has also been explored in recent years [197], [266], [267].
+It usually performs semi-supervised adaptation with the
+help of active learning [268], [269], model memorization rev-
+elation [270], and consistency and diversity learning [271].
+There is still a lot of space for improvement with a limited
+number of labeled target samples, e.g., by fine-tuning the
+current source-free adaptation frameworks, but this is not
+the focus of this survey.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+13
+6
+CONCLUSION
+In this paper, we provide a comprehensive review of recent
+progress in source-free unsupervised domain adaptation
+(SFUDA). We classify existing SFUDA studies into white-
+box and black-box groups, and each group is further catego-
+rized based on different learning strategies. The challenges
+of methods in each category and our insights are provided.
+We then compare white-box and black-box SFUDA meth-
+ods, discuss effective techniques for improving adaptation
+performance, and summarize commonly used datasets. We
+finally discuss promising future research directions. It is
+worth noting that the research topic of source-free unsu-
+pervised domain adaptation is still in its early stages, and
+we hope this survey can spark new ideas and attract more
+researchers to advance this high-impact research field.
+ACKNOWLEDGMENT
+This work was supported by NIH grants RF1AG073297 and
+R01MH108560.
+REFERENCES
+[1]
+A. Voulodimos, N. Doulamis, A. Doulamis, and E. Protopa-
+padakis, “Deep learning for computer vision: A brief review,”
+Computational Intelligence and Neuroscience, vol. 2018, 2018.
+[2]
+M. Hassaballah and A. I. Awad, Deep learning in computer vision:
+Principles and applications.
+CRC Press, 2020.
+[3]
+D. Shen, G. Wu, and H.-I. Suk, “Deep learning in medical image
+analysis,” Annual Review of Biomedical Engineering, vol. 19, p. 221,
+2017.
+[4]
+G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi,
+M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I.
+S´anchez, “A survey on deep learning in medical image analysis,”
+Medical Image Analysis, vol. 42, pp. 60–88, 2017.
+[5]
+D. W. Otter, J. R. Medina, and J. K. Kalita, “A survey of the
+usages of deep learning for natural language processing,” IEEE
+Transactions on Neural Networks and Learning Systems, vol. 32, no. 2,
+pp. 604–624, 2020.
+[6]
+T. Young, D. Hazarika, S. Poria, and E. Cambria, “Recent trends
+in deep learning based natural language processing,” IEEE Com-
+putational Intelligence Magazine, vol. 13, no. 3, pp. 55–75, 2018.
+[7]
+D. Li, Y. Yang, Y.-Z. Song, and T. M. Hospedales, “Deeper, broader
+and artier domain generalization,” in Proceedings of the IEEE
+International Conference on Computer Vision, 2017, pp. 5542–5550.
+[8]
+S. Sankaranarayanan, Y. Balaji, A. Jain, S. N. Lim, and R. Chel-
+lappa, “Learning from synthetic data: Addressing domain shift
+for semantic segmentation,” in Proceedings of the IEEE Conference
+on Computer Vision and Pattern Recognition, 2018, pp. 3752–3761.
+[9]
+K. Zhou, Z. Liu, Y. Qiao, T. Xiang, and C. C. Loy, “Domain
+generalization: A survey,” IEEE Transactions on Pattern Analysis
+and Machine Intelligence, 2022.
+[10]
+M. Wang and W. Deng, “Deep visual domain adaptation: A
+survey,” Neurocomputing, vol. 312, pp. 135–153, 2018.
+[11]
+H. Guan and M. Liu, “Domain adaptation for medical image
+analysis: A survey,” IEEE Transactions on Biomedical Engineering,
+vol. 69, no. 3, pp. 1173–1185, 2021.
+[12]
+J. Dong, Y. Cong, G. Sun, Z. Fang, and Z. Ding, “Where and how
+to transfer: Knowledge aggregation-induced transferability per-
+ception for unsupervised domain adaptation,” IEEE Transactions
+on Pattern Analysis and Machine Intelligence, 2021.
+[13]
+Y. Ganin and V. Lempitsky, “Unsupervised domain adaptation by
+backpropagation,” in International Conference on Machine Learning.
+PMLR, 2015, pp. 1180–1189.
+[14]
+K. Saito, K. Watanabe, Y. Ushiku, and T. Harada, “Maximum
+classifier discrepancy for unsupervised domain adaptation,” in
+Proceedings of the IEEE Conference on Computer Vision and Pattern
+Recognition, 2018, pp. 3723–3732.
+[15]
+Y. Fang, M. Wang, G. G. Potter, and M. Liu, “Unsupervised
+cross-domain functional MRI adaptation for automated major
+depressive disorder identification,” Medical Image Analysis, p.
+102707, 2022.
+[16]
+J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei,
+“ImageNet: A large-scale hierarchical image database,” in 2009
+IEEE Conference on Computer Vision and Pattern Recognition. IEEE,
+2009, pp. 248–255.
+[17]
+G. K. Nayak, K. R. Mopuri, S. Jain, and A. Chakraborty, “Min-
+ing data impressions from deep models as substitute for the
+unavailable training data,” IEEE Transactions on Pattern Analysis
+and Machine Intelligence, 2021.
+[18]
+J. N. Kundu, N. Venkat, R. V. Babu et al., “Universal source-free
+domain adaptation,” in Proceedings of the IEEE/CVF Conference on
+Computer Vision and Pattern Recognition, 2020, pp. 4544–4553.
+[19]
+J. Huang, D. Guan, A. Xiao, and S. Lu, “Model adaptation: His-
+torical contrastive learning for unsupervised domain adaptation
+without source data,” Advances in Neural Information Processing
+Systems, vol. 34, pp. 3635–3649, 2021.
+[20]
+S. Yang, J. van de Weijer, L. Herranz, S. Jui et al., “Exploiting the
+intrinsic neighborhood structure for source-free domain adapta-
+tion,” Advances in Neural Information Processing Systems, vol. 34,
+pp. 29 393–29 405, 2021.
+[21]
+Y. Liu, W. Zhang, and J. Wang, “Source-free domain adaptation
+for semantic segmentation,” in Proceedings of the IEEE/CVF Con-
+ference on Computer Vision and Pattern Recognition, 2021, pp. 1215–
+1224.
+[22]
+W. Chen, L. Lin, S. Yang, D. Xie, S. Pu, Y. Zhuang, and W. Ren,
+“Self-supervised noisy label learning for source-free unsuper-
+vised domain adaptation,” arXiv preprint arXiv:2102.11614, 2021.
+[23]
+C. Saltori, S. Lathuili´ere, N. Sebe, E. Ricci, and F. Galasso,
+“SF-UDA3D: Source-free unsupervised domain adaptation for
+LiDAR-based 3D object detection,” in 2020 International Confer-
+ence on 3D Vision (3DV).
+IEEE, 2020, pp. 771–780.
+[24]
+Y. Liu, Y. Chen, W. Dai, M. Gou, C.-T. Huang, and H. Xiong,
+“Source-free domain adaptation with contrastive domain align-
+ment and self-supervised exploration for face anti-spoofing,” in
+European Conference on Computer Vision.
+Springer, 2022, pp. 511–
+528.
+[25]
+Y. Liu, W. Zhang, J. Wang, and J. Wang, “Data-free knowledge
+transfer: A survey,” arXiv preprint arXiv:2112.15278, 2021.
+[26]
+C. Yang, X. Guo, Z. Chen, and Y. Yuan, “Source free domain
+adaptation for medical image segmentation with fourier style
+mining,” Medical Image Analysis, vol. 79, p. 102457, 2022.
+[27]
+Y. Hou and L. Zheng, “Source free domain adaptation with image
+translation,” arXiv preprint arXiv:2008.07514, 2020.
+[28]
+J. Hong, Y.-D. Zhang, and W. Chen, “Source-free unsupervised
+domain adaptation for cross-modality abdominal multi-organ
+segmentation,��� Knowledge-Based Systems, p. 109155, 2022.
+[29]
+Q. Tian, C. Ma, F.-Y. Zhang, S. Peng, and H. Xue, “Source-free
+unsupervised domain adaptation with sample transport learn-
+ing,” Journal of Computer Science and Technology, vol. 36, no. 3, pp.
+606–616, 2021.
+[30]
+Y. Ding, L. Sheng, J. Liang, A. Zheng, and R. He, “ProxyMix:
+Proxy-based mixup training with label refinery for source-free
+domain adaptation,” arXiv preprint arXiv:2205.14566, 2022.
+[31]
+M. Ye, J. Zhang, J. Ouyang, and D. Yuan, “Source data-free
+unsupervised domain adaptation for semantic segmentation,” in
+Proceedings of the 29th ACM International Conference on Multimedia,
+2021, pp. 2233–2242.
+[32]
+Y. Du, H. Yang, M. Chen, J. Jiang, H. Luo, and C. Wang, “Gen-
+eration, augmentation, and alignment: A pseudo-source domain
+based method for source-free domain adaptation,” arXiv preprint
+arXiv:2109.04015, 2021.
+[33]
+H. Yao, Y. Guo, and C. Yang, “Source-free unsupervised domain
+adaptation with surrogate data generation,” in Proceedings of
+NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods
+and Applications, 2021.
+[34]
+H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz,
+“Mixup: Beyond empirical risk minimization,” arXiv preprint
+arXiv:1710.09412, 2017.
+[35]
+S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein et al., “Dis-
+tributed optimization and statistical learning via the alternating
+direction method of multipliers,” Foundations and Trends® in
+Machine learning, vol. 3, no. 1, pp. 1–122, 2011.
+[36]
+V. K. Kurmi, V. K. Subramanian, and V. P. Namboodiri, “Domain
+impression: A source data free domain adaptation method,” in
+Proceedings of the IEEE/CVF Winter Conference on Applications of
+Computer Vision, 2021, pp. 615–625.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+14
+[37]
+Y. Hou and L. Zheng, “Visualizing adapted knowledge in domain
+transfer,” in Proceedings of the IEEE/CVF Conference on Computer
+Vision and Pattern Recognition, 2021, pp. 13 824–13 833.
+[38]
+R. Li, Q. Jiao, W. Cao, H.-S. Wong, and S. Wu, “Model adaptation:
+Unsupervised domain adaptation without source data,” in Pro-
+ceedings of the IEEE/CVF Conference on Computer Vision and Pattern
+Recognition, 2020, pp. 9641–9650.
+[39]
+Z. Qiu, Y. Zhang, H. Lin, S. Niu, Y. Liu, Q. Du, and M. Tan,
+“Source-free domain adaptation via avatar prototype generation
+and adaptation,” arXiv preprint arXiv:2106.15326, 2021.
+[40]
+J. Tian, J. Zhang, W. Li, and D. Xu, “VDM-DA: Virtual domain
+modeling for source data-free domain adaptation,” IEEE Transac-
+tions on Circuits and Systems for Video Technology, 2021.
+[41]
+N. Ding, Y. Xu, Y. Tang, C. Xu, Y. Wang, and D. Tao, “Source-free
+domain adaptation via distribution estimation,” in Proceedings of
+the IEEE/CVF Conference on Computer Vision and Pattern Recogni-
+tion, 2022, pp. 7212–7222.
+[42]
+S. Stan and M. Rostami, “Unsupervised model adaptation for
+continual semantic segmentation,” in Proceedings of the AAAI
+Conference on Artificial Intelligence, vol. 35, no. 3, 2021, pp. 2593–
+2601.
+[43]
+Stan, Serban and Rostami, Mohammad, “Privacy preserving do-
+main adaptation for semantic segmentation of medical images,”
+arXiv preprint arXiv:2101.00522, 2021.
+[44]
+W.-Y. Chen, Y.-C. Liu, Z. Kira, Y.-C. F. Wang, and J.-B.
+Huang, “A closer look at few-shot classification,” arXiv preprint
+arXiv:1904.04232, 2019.
+[45]
+G. Kang, L. Jiang, Y. Yang, and A. G. Hauptmann, “Contrastive
+adaptation network for unsupervised domain adaptation,” in
+Proceedings of the IEEE/CVF Conference on Computer Vision and
+Pattern Recognition, 2019, pp. 4893–4902.
+[46]
+C.-Y. Lee, T. Batra, M. H. Baig, and D. Ulbricht, “Sliced wasser-
+stein discrepancy for unsupervised domain adaptation,” in Pro-
+ceedings of the IEEE/CVF Conference on Computer Vision and Pattern
+Recognition, 2019, pp. 10 285–10 295.
+[47]
+D. Bang and H. Shim, “MGGAN: Solving mode collapse using
+manifold-guided training,” in Proceedings of the IEEE/CVF Inter-
+national Conference on Computer Vision, 2021, pp. 2347–2356.
+[48]
+A. Abusitta, O. A. Wahab, and B. C. Fung, “VirtualGAN: Re-
+ducing mode collapse in generative adversarial networks using
+virtual mapping,” in 2021 International Joint Conference on Neural
+Networks (IJCNN).
+IEEE, 2021, pp. 1–6.
+[49]
+X. Liu and Y. Yuan, “A source-free domain adaptive polyp detec-
+tion framework with style diversification flow,” IEEE Transactions
+on Medical Imaging, vol. 41, no. 7, pp. 1897–1908, 2022.
+[50]
+G. Yang, H. Tang, Z. Zhong, M. Ding, L. Shao, N. Sebe, and
+E. Ricci, “Transformer-based source-free domain adaptation,”
+arXiv preprint arXiv:2105.14138, 2021.
+[51]
+L. Xiong, M. Ye, D. Zhang, Y. Gan, X. Li, and Y. Zhu, “Source
+data-free domain adaptation of object detector through domain-
+specific perturbation,” International Journal of Intelligent Systems,
+vol. 36, no. 8, pp. 3746–3766, 2021.
+[52]
+X. Liu and S. Zhang, “Graph consistency based mean-teaching for
+unsupervised domain adaptive person re-identification,” arXiv
+preprint arXiv:2105.04776, 2021.
+[53]
+H. Yu, J. Huang, Y. Liu, Q. Zhu, M. Zhou, and F. Zhao, “Source-
+free domain adaptation for real-world image dehazing,” in Pro-
+ceedings of the 30th ACM International Conference on Multimedia,
+2022, pp. 6645–6654.
+[54]
+S. Tang, Y. Shi, Z. Ma, J. Li, J. Lyu, Q. Li, and J. Zhang, “Model
+adaptation through hypothesis transfer with gradual knowledge
+distillation,” in 2021 IEEE/RSJ International Conference on Intelli-
+gent Robots and Systems (IROS).
+IEEE, 2021, pp. 5679–5685.
+[55]
+V. VS, J. M. J. Valanarasu, and V. M. Patel, “Target and task
+specific source-free domain adaptive image segmentation,” arXiv
+preprint arXiv:2203.15792, 2022.
+[56]
+A. Tarvainen and H. Valpola, “Mean teachers are better role
+models: Weight-averaged consistency targets improve semi-
+supervised deep learning results,” Advances in Neural Information
+Processing Systems, vol. 30, 2017.
+[57]
+A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai,
+T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly
+et al., “An image is worth 16×16 words: Transformers for image
+recognition at scale,” arXiv preprint arXiv:2010.11929, 2020.
+[58]
+M. Ishii and M. Sugiyama, “Source-free domain adaptation via
+distributional alignment by matching batch normalization statis-
+tics,” arXiv preprint arXiv:2101.10842, 2021.
+[59]
+X. Liu, F. Xing, C. Yang, G. El Fakhri, and J. Woo, “Adapting
+off-the-shelf source segmenter for target medical image segmen-
+tation,” in International Conference on Medical Image Computing and
+Computer-Assisted Intervention.
+Springer, 2021, pp. 549–559.
+[60]
+S. Paul, A. Khurana, and G. Aggarwal, “Unsupervised adap-
+tation of semantic segmentation models without source data,”
+arXiv preprint arXiv:2112.02359, 2021.
+[61]
+J. Fan, H. Zhu, X. Jiang, L. Meng, C. Chen, C. Fu, H. Yu, C. Dai,
+and W. Chen, “Unsupervised domain adaptation by statistics
+alignment for deep sleep staging networks,” IEEE Transactions on
+Neural Systems and Rehabilitation Engineering, vol. 30, pp. 205–216,
+2022.
+[62]
+C. Eastwood, I. Mason, C. K. Williams, and B. Sch¨olkopf, “Source-
+free adaptation to measurement shift via bottom-up feature
+restoration,” arXiv preprint arXiv:2107.05446, 2021.
+[63]
+D. Zhang, M. Ye, L. Xiong, S. Li, and X. Li, “Source-style trans-
+ferred mean teacher for source-data free object detection,” in
+ACM Multimedia Asia, 2021, pp. 1–8.
+[64]
+M. Klingner, J.-A. Term¨ohlen, J. Ritterbach, and T. Fingscheidt,
+“Unsupervised batchnorm adaptation (UBNA): A domain adap-
+tation method for semantic segmentation without using source
+domain representations,” in Proceedings of the IEEE/CVF Winter
+Conference on Applications of Computer Vision, 2022, pp. 210–220.
+[65]
+W.-G. Chang, T. You, S. Seo, S. Kwak, and B. Han, “Domain-
+specific batch normalization for unsupervised domain adapta-
+tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision
+and Pattern Recognition, 2019, pp. 7354–7362.
+[66]
+Y. Li, N. Wang, J. Shi, J. Liu, and X. Hou, “Revisiting batch
+normalization for practical domain adaptation,” arXiv preprint
+arXiv:1603.04779, 2016.
+[67]
+S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
+network training by reducing internal covariate shift,” in Interna-
+tional Conference on Machine Learning.
+PMLR, 2015, pp. 448–456.
+[68]
+D. Ulyanov, A. Vedaldi, and V. Lempitsky, “Instance normaliza-
+tion: The missing ingredient for fast stylization,” arXiv preprint
+arXiv:1607.08022, 2016.
+[69]
+H. Xia, H. Zhao, and Z. Ding, “Adaptive adversarial network for
+source-free domain adaptation,” in Proceedings of the IEEE/CVF
+International Conference on Computer Vision, 2021, pp. 9010–9019.
+[70]
+R. Wang, Z. Wu, Z. Weng, J. Chen, G.-J. Qi, and Y.-G. Jiang,
+“Cross-domain contrastive learning for unsupervised domain
+adaptation,” IEEE Transactions on Multimedia, 2022.
+[71]
+P. Agarwal, D. P. Paudel, J.-N. Zaech, and L. Van Gool, “Un-
+supervised robust domain adaptation without source data,” in
+Proceedings of the IEEE/CVF Winter Conference on Applications of
+Computer Vision, 2022, pp. 2009–2018.
+[72]
+X. Zhao, R. Stanislawski, P. Gardoni, M. Sulowicz, A. Glowacz,
+G. Krolczyk, and Z. Li, “Adaptive contrastive learning with label
+consistency for source data free unsupervised domain adapta-
+tion,” Sensors, vol. 22, no. 11, p. 4238, 2022.
+[73]
+K. Sohn, “Improved deep metric learning with multi-class N-pair
+loss objective,” Advances in Neural Information Processing Systems,
+vol. 29, 2016.
+[74]
+J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt,
+J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscher et al., “A
+survey of uncertainty in deep neural networks,” arXiv preprint
+arXiv:2107.03342, 2021.
+[75]
+F. Fleuret et al., “Uncertainty reduction for model adaptation in
+semantic segmentation,” in Proceedings of the IEEE/CVF Conference
+on Computer Vision and Pattern Recognition, 2021, pp. 9613–9623.
+[76]
+J. Lee and G. Lee, “Feature alignment by uncertainty and self-
+training for source-free unsupervised domain adaptation,” arXiv
+preprint arXiv:2208.14888, 2022.
+[77]
+C. Chen, Q. Liu, Y. Jin, Q. Dou, and P.-A. Heng, “Source-free
+domain adaptive fundus image segmentation with denoised
+pseudo-labeling,” in International Conference on Medical Image
+Computing and Computer-Assisted Intervention.
+Springer, 2021,
+pp. 225–235.
+[78]
+Z. Xu, D. Lu, Y. Wang, J. Luo, D. Wei, Y. Zheng, and R. K.-y.
+Tong, “Denoising for relaxing: Unsupervised domain adaptive
+fundus image segmentation without source data,” in International
+Conference on Medical Image Computing and Computer-Assisted In-
+tervention.
+Springer, 2022, pp. 214–224.
+[79]
+D. Hegde, V. Sindagi, V. Kilic, A. B. Cooper, M. Foster, and
+V. Patel, “Uncertainty-aware mean teacher for source-free unsu-
+pervised domain adaptive 3D object detection,” arXiv preprint
+arXiv:2109.14651, 2021.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+15
+[80]
+S. Roy, M. Trapp, A. Pilzer, J. Kannala, N. Sebe, E. Ricci, and
+A. Solin, “Uncertainty-guided source-free domain adaptation,”
+in European Conference on Computer Vision.
+Springer, 2022, pp.
+537–555.
+[81]
+J. Pei, Z. Jiang, A. Men, L. Chen, Y. Liu, and Q. Chen,
+“Uncertainty-induced
+transferability
+representation
+for
+source-free unsupervised domain adaptation,” arXiv preprint
+arXiv:2208.13986, 2022.
+[82]
+X. Li, W. Chen, D. Xie, S. Yang, P. Yuan, S. Pu, and Y. Zhuang,
+“A free lunch for unsupervised domain adaptive object detection
+without source data,” in Proceedings of the AAAI Conference on
+Artificial Intelligence, vol. 35, no. 10, 2021, pp. 8474–8481.
+[83]
+N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
+R. Salakhutdinov, “Dropout: A simple way to prevent neural net-
+works from overfitting,” The Journal of Machine Learning Research,
+vol. 15, no. 1, pp. 1929–1958, 2014.
+[84]
+Y. Gal and Z. Ghahramani, “Dropout as a bayesian approxi-
+mation: Representing model uncertainty in deep learning,” in
+International Conference on Machine Learning.
+PMLR, 2016, pp.
+1050–1059.
+[85]
+A. Kendall and Y. Gal, “What uncertainties do we need in
+bayesian deep learning for computer vision?” Advances in Neural
+Information Processing Systems, vol. 30, 2017.
+[86]
+Z. Xu, D. Lu, Y. Wang, J. Luo, J. Jayender, K. Ma, Y. Zheng, and
+X. Li, “Noisy labels are treasure: Mean-teacher-assisted confident
+learning for hepatic vessel segmentation,” in International Confer-
+ence on Medical Image Computing and Computer-Assisted Interven-
+tion.
+Springer, 2021, pp. 3–13.
+[87]
+C. Northcutt, L. Jiang, and I. Chuang, “Confident learning:
+Estimating uncertainty in dataset labels,” Journal of Artificial
+Intelligence Research, vol. 70, pp. 1373–1411, 2021.
+[88]
+Y. Gal, J. Hron, and A. Kendall, “Concrete dropout,” Advances in
+Neural Information Processing Systems, vol. 30, 2017.
+[89]
+L. Tierney and J. B. Kadane, “Accurate approximations for pos-
+terior moments and marginal densities,” Journal of the American
+Statistical Association, vol. 81, no. 393, pp. 82–86, 1986.
+[90]
+D. J. MacKay, D. J. Mac Kay et al., Information theory, inference and
+learning algorithms.
+Cambridge University Press, 2003.
+[91]
+A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Op-
+timal speed and accuracy of object detection,” arXiv preprint
+arXiv:2004.10934, 2020.
+[92]
+S. Yang, Y. Wang, J. van de Weijer, L. Herranz, and S. Jui,
+“Generalized source-free domain adaptation,” in Proceedings of
+the IEEE/CVF International Conference on Computer Vision, 2021,
+pp. 8978–8987.
+[93]
+S. Tang, Y. Yang, Z. Ma, N. Hendrich, F. Zeng, S. S. Ge, C. Zhang,
+and J. Zhang, “Nearest neighborhood-based deep clustering
+for source data-absent unsupervised domain adaptation,” arXiv
+preprint arXiv:2107.12585, 2021.
+[94]
+J. Liang, D. Hu, and J. Feng, “Do we really need to access the
+source data? Source hypothesis transfer for unsupervised do-
+main adaptation,” in International Conference on Machine Learning.
+PMLR, 2020, pp. 6028–6039.
+[95]
+J. Liang, D. Hu, Y. Wang, R. He, and J. Feng, “Source data-absent
+unsupervised domain adaptation through hypothesis transfer
+and labeling transfer,” IEEE Transactions on Pattern Analysis and
+Machine Intelligence, 2021.
+[96]
+J. Lee, D. Jung, J. Yim, and S. Yoon, “Confidence score for source-
+free unsupervised domain adaptation,” in International Conference
+on Machine Learning.
+PMLR, 2022, pp. 12 365–12 377.
+[97]
+S. Yang, Y. Wang, K. Wang, S. Jui, and J. van de Weijer, “Attract-
+ing and dispersing: A simple approach for source-free domain
+adaptation.” CoRR, 2022.
+[98]
+W. Li, M. Cao, and S. Chen, “Jacobian norm for unsupervised
+source-free domain adaptation,” arXiv preprint arXiv:2204.03467,
+2022.
+[99]
+F. G. Ashby, W. T. Maddox et al., “Human category learning,”
+Annual Review of Psychology, vol. 56, no. 1, pp. 149–178, 2005.
+[100] J. Li, Z. Du, L. Zhu, Z. Ding, K. Lu, and H. T. Shen, “Divergence-
+agnostic unsupervised domain adaptation by adversarial at-
+tacks,” IEEE Transactions on Pattern Analysis and Machine Intelli-
+gence, 2021.
+[101] F. Wang, Z. Han, Y. Gong, and Y. Yin, “Exploring domain-
+invariant parameters for source free domain adaptation,” in
+Proceedings of the IEEE/CVF Conference on Computer Vision and
+Pattern Recognition, 2022, pp. 7151–7160.
+[102] J. Liang, R. He, Z. Sun, and T. Tan, “Distant supervised centroid
+shift: A simple and efficient approach to visual domain adapta-
+tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision
+and Pattern Recognition, 2019, pp. 2975–2984.
+[103] S. Yang, Y. Wang, J. van de Weijer, L. Herranz, and S. Jui, “Casting
+a BAIT for offline and online source-free domain adaptation,”
+arXiv preprint arXiv:2010.12427, 2020.
+[104] Z. Ke, D. Wang, Q. Yan, J. Ren, and R. W. Lau, “Dual student:
+Breaking the limits of the teacher in semi-supervised learning,”
+in Proceedings of the IEEE/CVF International Conference on Computer
+Vision, 2019, pp. 6728–6736.
+[105] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple
+framework for contrastive learning of visual representations,” in
+International Conference on Machine Learning.
+PMLR, 2020, pp.
+1597–1607.
+[106] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum
+contrast for unsupervised visual representation learning,” in
+Proceedings of the IEEE/CVF Conference on Computer Vision and
+Pattern Recognition, 2020, pp. 9729–9738.
+[107] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple
+framework for contrastive learning of visual representations,” in
+International Conference on Machine Learning.
+PMLR, 2020, pp.
+1597–1607.
+[108] W. Chen, S. Pu, D. Xie, S. Yang, Y. Guo, and L. Lin, “Unsuper-
+vised image classification for deep representation learning,” in
+European Conference on Computer Vision.
+Springer, 2020, pp. 430–
+446.
+[109] J. Liang, D. Hu, J. Feng, and R. He, “DINE: Domain adaptation
+from single and multiple black-box predictors,” in Proceedings of
+the IEEE/CVF Conference on Computer Vision and Pattern Recogni-
+tion, 2022, pp. 8003–8013.
+[110] J. Liang, D. Hu, R. He, and J. Feng, “Distill and fine-tune: Effec-
+tive adaptation from a black-box source model,” arXiv preprint
+arXiv:2104.01539, 2021.
+[111] X. Liu, C. Yoo, F. Xing, C.-C. J. Kuo, G. El Fakhri, J.-W. Kang, and
+J. Woo, “Unsupervised black-box model domain adaptation for
+brain tumor segmentation,” Frontiers in Neuroscience, p. 341, 2022.
+[112] Liu, Xiaofeng and Yoo, Chaehwa and Xing, Fangxu and Kuo, C-C
+Jay and El Fakhri, Georges and Kang, Je-Won and Woo, Jonghye,
+“Unsupervised domain adaptation for segmentation with black-
+box source model,” in Medical Imaging 2022: Image Processing, vol.
+12032.
+SPIE, 2022, pp. 255–260.
+[113] Y. Xu, J. Yang, M. Wu, X. Li, L. Xie, and Z. Chen, “EXTERN:
+Leveraging endo-temporal regularization for black-box video
+domain adaptation,” arXiv preprint arXiv:2208.05187, 2022.
+[114] Q. Peng, Z. Ding, L. Lyu, L. Sun, and C. Chen, “Toward better
+target representation for source-free and black-box domain adap-
+tation,” arXiv preprint arXiv:2208.10531, 2022.
+[115] S. Laine and T. Aila, “Temporal ensembling for semi-supervised
+learning,” arXiv preprint arXiv:1610.02242, 2016.
+[116] K. Kim, B. Ji, D. Yoon, and S. Hwang, “Self-knowledge distillation
+with progressive refinement of targets,” in Proceedings of the
+IEEE/CVF International Conference on Computer Vision, 2021, pp.
+6567–6576.
+[117] T. Yang, S. Zhu, C. Chen, S. Yan, M. Zhang, and A. Willis, “Mu-
+tualNet: Adaptive convnet via mutual learning from network
+width and resolution,” in European Conference on Computer Vision.
+Springer, 2020, pp. 299–315.
+[118] H. Zhang, Y. Zhang, K. Jia, and L. Zhang, “Unsupervised do-
+main adaptation of black-box source models,” arXiv preprint
+arXiv:2101.02839, 2021.
+[119] X. Luo, W. Chen, Y. Tan, C. Li, Y. He, and X. Jia, “Exploiting
+negative learning for implicit pseudo label rectification in source-
+free domain adaptive semantic segmentation,” arXiv preprint
+arXiv:2106.12123, 2021.
+[120] D. Arpit, S. Jastrzebski, N. Ballas, D. Krueger, E. Bengio, M. S.
+Kanwal, T. Maharaj, A. Fischer, A. Courville, Y. Bengio et al., “A
+closer look at memorization in deep networks,” in International
+Conference on Machine Learning.
+PMLR, 2017, pp. 233–242.
+[121] J. Yang, X. Peng, K. Wang, Z. Zhu, J. Feng, L. Xie, and Y. You,
+“Divide to adapt: Mitigating confirmation bias for domain adap-
+tation of black-box predictors,” arXiv preprint arXiv:2205.14467,
+2022.
+[122] E. Arazo, D. Ortego, P. Albert, N. O’Connor, and K. McGuinness,
+“Unsupervised label noise modeling and loss correction,” in
+International Conference on Machine Learning.
+PMLR, 2019, pp.
+312–321.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+16
+[123] D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver,
+and C. A. Raffel, “MixMatch: A holistic approach to semi-
+supervised learning,” Advances in Neural Information Processing
+Systems, vol. 32, 2019.
+[124] H.-W. Yeh, B. Yang, P. C. Yuen, and T. Harada, “SoFA: Source-
+data-free feature alignment for unsupervised domain adapta-
+tion,” in Proceedings of the IEEE/CVF Winter Conference on Applica-
+tions of Computer Vision, 2021, pp. 474–483.
+[125] B. Yang, H.-W. Yeh, T. Harada, and P. C. Yuen, “Model-induced
+generalization error bound for information-theoretic representa-
+tion learning in source-data-free unsupervised domain adapta-
+tion,” IEEE Transactions on Image Processing, vol. 31, pp. 419–432,
+2021.
+[126] W. Zhang and D. Wu, “Lightweight source-free transfer for
+privacy-preserving motor imagery classification,” IEEE Transac-
+tions on Cognitive and Developmental Systems, 2022.
+[127] Zhang, Wen and Wu, Dongrui, “Discriminative joint probability
+maximum mean discrepancy (DJP-MMD) for domain adapta-
+tion,” in 2020 International Joint Conference on Neural Networks
+(IJCNN).
+IEEE, 2020, pp. 1–8.
+[128] X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang,
+“Moment matching for multi-source domain adaptation,” in
+Proceedings of the IEEE/CVF International Conference on Computer
+Vision, 2019, pp. 1406–1415.
+[129] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
+learning applied to document recognition,” Proceedings of the
+IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
+[130] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y.
+Ng, “Reading digits in natural images with unsupervised feature
+learning,” 2011.
+[131] J. J. Hull, “A database for handwritten text recognition research,”
+IEEE Transactions on Pattern Analysis and Machine Intelligence,
+vol. 16, no. 5, pp. 550–554, 1994.
+[132] S. R. Richter, V. Vineet, S. Roth, and V. Koltun, “Playing for data:
+Ground truth from computer games,” in European Conference on
+Computer Vision.
+Springer, 2016, pp. 102–118.
+[133] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler,
+R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes
+dataset for semantic urban scene understanding,” in Proceedings
+of the IEEE Conference on Computer Vision and Pattern Recognition,
+2016, pp. 3213–3223.
+[134] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M. Lopez,
+“The synthia dataset: A large collection of synthetic images for
+semantic segmentation of urban scenes,” in Proceedings of the IEEE
+Conference on Computer Vision and Pattern Recognition, 2016, pp.
+3234–3243.
+[135] Y.-H. Chen, W.-Y. Chen, Y.-T. Chen, B.-C. Tsai, Y.-C. Frank Wang,
+and M. Sun, “No more discrimination: Cross city adaptation of
+road scene segmenters,” in Proceedings of the IEEE International
+Conference on Computer Vision, 2017, pp. 1992–2001.
+[136] K. Saenko, B. Kulis, M. Fritz, and T. Darrell, “Adapting visual
+category models to new domains,” in European Conference on
+Computer Vision.
+Springer, 2010, pp. 213–226.
+[137] H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Pan-
+chanathan, “Deep hashing network for unsupervised domain
+adaptation,” in Proceedings of the IEEE Conference on Computer
+Vision and Pattern Recognition, 2017, pp. 5018–5027.
+[138] X. Peng, B. Usman, N. Kaushik, J. Hoffman, D. Wang, and
+K. Saenko, “VisDA: The visual domain adaptation challenge,”
+arXiv preprint arXiv:1710.06924, 2017.
+[139] B. Gong, Y. Shi, F. Sha, and K. Grauman, “Geodesic flow kernel
+for unsupervised domain adaptation,” in 2012 IEEE Conference on
+Computer Vision and Pattern Recognition.
+IEEE, 2012, pp. 2066–
+2073.
+[140] B. Caputo, H. M¨uller, J. Martinez-Gomez, M. Villegas, B. Acar,
+N. Patricia, N. Marvasti, S. ¨Usk¨udarlı, R. Paredes, M. Cazorla
+et al., “ImageCLEF 2014: Overview and analysis of the results,”
+in International Conference of the Cross-Language Evaluation Forum
+for European Languages.
+Springer, 2014, pp. 192–211.
+[141] G. Griffin, A. Holub, and P. Perona, “Caltech-256 object category
+dataset,” 2007.
+[142] M. Everingham and J. Winn, “The PASCAL visual object classes
+challenge 2007 (VOC2007) development kit,” International Journal
+of Computer Vision, vol. 88, no. 2, pp. 303–338, 2010.
+[143] A. Bergamo and L. Torresani, “Exploiting weakly-labeled web
+images to improve object classification: A domain adaptation
+approach,” Advances in Neural Information Processing Systems,
+vol. 23, 2010.
+[144] X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang,
+“Moment matching for multi-source domain adaptation,” in
+Proceedings of the IEEE/CVF International Conference on Computer
+Vision, 2019, pp. 1406–1415.
+[145] K. Zhou, Y. Yang, Y. Qiao, and T. Xiang, “Domain adaptive
+ensemble learning,” IEEE Transactions on Image Processing, vol. 30,
+pp. 8008–8018, 2021.
+[146] C. Qin, H. You, L. Wang, C.-C. J. Kuo, and Y. Fu, “PointDAN:
+A multi-scale 3D domain adaption network for point cloud
+representation,” Advances in Neural Information Processing Systems,
+vol. 32, 2019.
+[147] Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao,
+“3D ShapeNets: A deep representation for volumetric shapes,” in
+Proceedings of the IEEE Conference on Computer Vision and Pattern
+Recognition, 2015, pp. 1912–1920.
+[148] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang,
+Z. Li, S. Savarese, M. Savva, S. Song, H. Su et al., “ShapeNet:
+An
+information-rich
+3D
+model
+repository,”
+arXiv
+preprint
+arXiv:1512.03012, 2015.
+[149] A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and
+M. Nießner, “ScanNet: Richly-annotated 3D reconstructions of
+indoor scenes,” in Proceedings of the IEEE Conference on Computer
+Vision and Pattern Recognition, 2017, pp. 5828–5839.
+[150] I. Chingovska, A. Anjos, and S. Marcel, “On the effectiveness
+of local binary patterns in face anti-spoofing,” in 2012 BIOSIG-
+Proceedings of the International Conference of Biometrics Special Inter-
+est Group (BIOSIG).
+IEEE, 2012, pp. 1–7.
+[151] Z. Boulkenafet, J. Komulainen, L. Li, X. Feng, and A. Hadid,
+“OULU-NPU: A mobile face presentation attack database with
+real-world variations,” in 2017 12th IEEE International Conference
+on Automatic Face & Gesture Recognition (FG 2017).
+IEEE, 2017,
+pp. 612–618.
+[152] Z. Zhang, J. Yan, S. Liu, Z. Lei, D. Yi, and S. Z. Li, “A face
+antispoofing database with diverse attacks,” in 2012 5th IAPR
+International Conference on Biometrics (ICB).
+IEEE, 2012, pp. 26–
+31.
+[153] D. Wen, H. Han, and A. K. Jain, “Face spoof detection with image
+distortion analysis,” IEEE Transactions on Information Forensics and
+Security, vol. 10, no. 4, pp. 746–761, 2015.
+[154] P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V. Patnaik,
+P. Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine et al., “Scalability in
+perception for autonomous driving: Waymo open dataset,” in
+Proceedings of the IEEE/CVF Conference on Computer Vision and
+Pattern Recognition, 2020, pp. 2446–2454.
+[155] A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for au-
+tonomous driving? The KITTI vision benchmark suite,” in 2012
+IEEE Conference on Computer Vision and Pattern Recognition. IEEE,
+2012, pp. 3354–3361.
+[156] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu,
+A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuscenes: A
+multimodal dataset for autonomous driving,” in Proceedings of the
+IEEE/CVF Conference on Computer Vision and Pattern Recognition,
+2020, pp. 11 621–11 631.
+[157] M.-H. Chen, Z. Kira, G. AlRegib, J. Yoo, R. Chen, and J. Zheng,
+“Temporal attentive alignment for large-scale video domain
+adaptation,” in Proceedings of the IEEE/CVF International Confer-
+ence on Computer Vision, 2019, pp. 6321–6330.
+[158] K. Soomro, A. R. Zamir, and M. Shah, “UCF101: A dataset of 101
+human actions classes from videos in the wild,” arXiv preprint
+arXiv:1212.0402, 2012.
+[159] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre,
+“HMDB: A large video database for human motion recognition,”
+in 2011 International Conference on Computer Vision.
+IEEE, 2011,
+pp. 2556–2563.
+[160] Y. Xu, J. Yang, H. Cao, K. Wu, M. Wu, R. Zhao, and Z. Chen,
+“Multi-source video domain adaptation with temporal attentive
+moment alignment,” arXiv preprint arXiv:2109.09964, 2021.
+[161] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
+L. Fei-Fei, “Large-scale video classification with convolutional
+neural networks,” in Proceedings of the IEEE Conference on Com-
+puter Vision and Pattern Recognition, 2014, pp. 1725–1732.
+[162] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya-
+narasimhan, F. Viola, T. Green, T. Back, P. Natsev et al., “The kinet-
+ics human action video dataset,” arXiv preprint arXiv:1705.06950,
+2017.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+17
+[163] Y. Xu, J. Yang, H. Cao, K. Mao, J. Yin, and S. See, “ARID: A new
+dataset for recognizing action in the dark,” in International Work-
+shop on Deep Learning for Human Activity Recognition.
+Springer,
+2021, pp. 70–84.
+[164] M. Monfort, A. Andonian, B. Zhou, K. Ramakrishnan, S. A.
+Bargal, T. Yan, L. Brown, Q. Fan, D. Gutfreund, C. Vondrick
+et al., “Moments in time dataset: One million videos for event
+understanding,” IEEE Transactions on Pattern Analysis and Machine
+Intelligence, vol. 42, no. 2, pp. 502–508, 2019.
+[165] B. Moiseev, A. Konev, A. Chigorin, and A. Konushin, “Evalua-
+tion of traffic sign recognition methods trained on synthetically
+generated data,” in International Conference on Advanced Concepts
+for Intelligent Vision Systems.
+Springer, 2013, pp. 576–583.
+[166] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel, “The german
+traffic sign recognition benchmark: A multi-class classification
+competition,” in The 2011 International Joint Conference on Neural
+Networks.
+IEEE, 2011, pp. 1453–1460.
+[167] C. Fang, Y. Xu, and D. N. Rockmore, “Unbiased metric learning:
+On the utilization of multiple datasets and web images for
+softening bias,” in Proceedings of the IEEE International Conference
+on Computer Vision, 2013, pp. 1657–1664.
+[168] L. Fei-Fei, R. Fergus, and P. Perona, “One-shot learning of object
+categories,” IEEE Transactions on Pattern Analysis and Machine
+Intelligence, vol. 28, no. 4, pp. 594–611, 2006.
+[169] B. C. Russell, A. Torralba, K. P. Murphy, and W. T. Freeman,
+“LabelMe: A database and web-based tool for image annotation,”
+International Journal of Computer Vision, vol. 77, no. 1, pp. 157–173,
+2008.
+[170] M. J. Choi, J. J. Lim, A. Torralba, and A. S. Willsky, “Exploiting
+hierarchical context on a large database of object categories,”
+in 2010 IEEE Computer Society Conference on Computer Vision and
+Pattern Recognition.
+IEEE, 2010, pp. 129–136.
+[171] B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani,
+J. Kirby, Y. Burren, N. Porz, J. Slotboom, R. Wiest et al., “The mul-
+timodal brain tumor image segmentation benchmark (BRATS),”
+IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993–
+2024, 2014.
+[172] X. Zhuang and J. Shen, “Multi-scale patch and multi-modality
+atlases for whole heart segmentation of MRI,” Medical Image
+Analysis, vol. 31, pp. 77–87, 2016.
+[173] Y. Li, R. Dan, S. Wang, Y. Cao, X. Luo, C. Tan, G. Jia, H. Zhou,
+Y. Wang, and L. Wang, “Source-free domain adaptation for
+multi-site and lifespan brain skull stripping,” arXiv preprint
+arXiv:2203.04299, 2022.
+[174] S. F. Eskildsen, P. Coup´e, V. Fonov, J. V. Manj´on, K. K. Leung,
+N. Guizard, S. N. Wassef, L. R. Østergaard, D. L. Collins, A. D. N.
+Initiative et al., “BEaST: Brain extraction based on nonlocal seg-
+mentation technique,” NeuroImage, vol. 59, no. 3, pp. 2362–2373,
+2012.
+[175] C. R. Jack Jr, M. A. Bernstein, N. C. Fox, P. Thompson, G. Alexan-
+der, D. Harvey, B. Borowski, P. J. Britson, J. L. Whitwell, C. Ward
+et al., “The Alzheimer’s disease neuroimaging initiative (ADNI):
+MRI methods,” Journal of Magnetic Resonance Imaging: An Offi-
+cial Journal of the International Society for Magnetic Resonance in
+Medicine, vol. 27, no. 4, pp. 685–691, 2008.
+[176] A. Makropoulos, E. C. Robinson, A. Schuh, R. Wright, S. Fitzgib-
+bon, J. Bozek, S. J. Counsell, J. Steinweg, K. Vecchiato, J. Passerat-
+Palmbach et al., “The developing human connectome project: A
+minimal processing pipeline for neonatal cortical surface recon-
+struction,” NeuroImage, vol. 173, pp. 88–112, 2018.
+[177] J.
+Bernal,
+F.
+J.
+S´anchez,
+G.
+Fern´andez-Esparrach,
+D.
+Gil,
+C. Rodr´ıguez, and F. Vilari˜no, “WM-DOVA maps for accurate
+polyp highlighting in colonoscopy: Validation vs. saliency maps
+from physicians,” Computerized Medical Imaging and Graphics,
+vol. 43, pp. 99–111, 2015.
+[178] T.-H. Hoang, H.-D. Nguyen, V.-A. Nguyen, T.-A. Nguyen, V.-T.
+Nguyen, and M.-T. Tran, “Enhancing endoscopic image classi-
+fication with symptom localization and data augmentation,” in
+Proceedings of the 27th ACM International Conference on Multimedia,
+2019, pp. 2578–2582.
+[179] J. Silva, A. Histace, O. Romain, X. Dray, and B. Granado, “Toward
+embedded detection of polyps in WCE images for early diagnosis
+of colorectal cancer,” International Journal of Computer Assisted
+Radiology and Surgery, vol. 9, no. 2, pp. 283–293, 2014.
+[180] D. V´azquez, J. Bernal, F. J. S´anchez, G. Fern´andez-Esparrach,
+A. M. L´opez, A. Romero, M. Drozdzal, and A. Courville, “A
+benchmark for endoluminal scene segmentation of colonoscopy
+images,” Journal of Healthcare Engineering, vol. 2017, 2017.
+[181] M. Tangermann, K.-R. M¨uller, A. Aertsen, N.
+Birbaumer,
+C. Braun, C. Brunner, R. Leeb, C. Mehring, K. J. Miller,
+G. Mueller-Putz et al., “Review of the BCI competition IV,”
+Frontiers in Neuroscience, p. 55, 2012.
+[182] J. Faller, C. Vidaurre, T. Solis-Escalante, C. Neuper, and R. Scherer,
+“Autocalibration and recurrent adaptation: Towards a plug and
+play online ERD-BCI,” IEEE Transactions on Neural Systems and
+Rehabilitation Engineering, vol. 20, no. 3, pp. 313–319, 2012.
+[183] V. Jayaram and A. Barachant, “MOABB: Trustworthy algorithm
+benchmarking for BCIs,” Journal of Neural Engineering, vol. 15,
+no. 6, p. 066011, 2018.
+[184] N. Bloch, A. Madabhushi, H. Huisman, J. Freymann, J. Kirby,
+M. Grauer, A. Enquobahrie, C. Jaffe, L. Clarke, and K. Farahani,
+“NCI-ISBI 2013 challenge: Automated segmentation of prostate
+structures,” The Cancer Imaging Archive, vol. 370, no. 6, p. 5, 2015.
+[185] G. Litjens, R. Toth, W. van de Ven, C. Hoeks, S. Kerkstra, B. van
+Ginneken, G. Vincent, G. Guillard, N. Birbeck, J. Zhang et al.,
+“Evaluation of prostate segmentation algorithms for MRI: The
+PROMISE12 challenge,” Medical Image Analysis, vol. 18, no. 2, pp.
+359–373, 2014.
+[186] J. I. Orlando, H. Fu, J. B. Breda, K. van Keer, D. R. Bathula,
+A. Diaz-Pinto, R. Fang, P.-A. Heng, J. Kim, J. Lee et al., “Refuge
+challenge: A unified framework for evaluating automated meth-
+ods for glaucoma assessment from fundus photographs,” Medical
+Image Analysis, vol. 59, p. 101570, 2020.
+[187] F. Fumero, S. Alay´on, J. L. Sanchez, J. Sigut, and M. Gonzalez-
+Hernandez, “RIM-ONE: An open retinal image database for
+optic nerve evaluation,” in 2011 24th International Symposium on
+Computer-based Medical Systems (CBMS).
+IEEE, 2011, pp. 1–6.
+[188] J. Sivaswamy, S. Krishnadas, A. Chakravarty, G. Joshi, A. S.
+Tabish et al., “A comprehensive retinal image dataset for the
+assessment of glaucoma from the optic nerve head analysis,” JSM
+Biomedical Imaging Data Papers, vol. 2, no. 1, p. 1004, 2015.
+[189] A. Di Martino, C.-G. Yan, Q. Li, E. Denio, F. X. Castel-
+lanos, K. Alaerts, J. S. Anderson, M. Assaf, S. Y. Bookheimer,
+M. Dapretto et al., “The autism brain imaging data exchange: To-
+wards a large-scale evaluation of the intrinsic brain architecture
+in autism,” Molecular Psychiatry, vol. 19, no. 6, pp. 659–667, 2014.
+[190] H. Yin, P. Molchanov, J. M. Alvarez, Z. Li, A. Mallya, D. Hoiem,
+N. K. Jha, and J. Kautz, “Dreaming to distill: Data-free knowl-
+edge transfer via DeepInversion,” in Proceedings of the IEEE/CVF
+Conference on Computer Vision and Pattern Recognition, 2020, pp.
+8715–8724.
+[191] M. Nasr, R. Shokri, and A. Houmansadr, “Comprehensive pri-
+vacy analysis of deep learning: Passive and active white-box
+inference attacks against centralized and federated learning,” in
+2019 IEEE Symposium on Security and Privacy (SP).
+IEEE, 2019,
+pp. 739–753.
+[192] H. Hu, Z. Salcic, L. Sun, G. Dobbie, P. S. Yu, and X. Zhang,
+“Membership inference attacks on machine learning: A survey,”
+ACM Computing Surveys (CSUR), vol. 54, no. 11s, pp. 1–37, 2022.
+[193] S. M. Ahmed, D. S. Raychaudhuri, S. Paul, S. Oymak, and A. K.
+Roy-Chowdhury, “Unsupervised multi-source domain adapta-
+tion without access to source data,” in Proceedings of the IEEE/CVF
+Conference on Computer Vision and Pattern Recognition, 2021, pp.
+10 103–10 112.
+[194] Y. Grandvalet and Y. Bengio, “Semi-supervised learning by en-
+tropy minimization,” Advances in Neural Information Processing
+Systems, vol. 17, 2004.
+[195] M. Bateson, H. Kervadec, J. Dolz, H. Lombaert, and I. B. Ayed,
+“Source-free domain adaptation for image segmentation,” Medi-
+cal Image Analysis, vol. 82, p. 102617, 2022.
+[196] J. Liu, X. Li, S. An, and Z. Chen, “Source-free unsupervised
+domain adaptation for blind image quality assessment,” arXiv
+preprint arXiv:2207.08124, 2022.
+[197] D. Kothandaraman, R. Chandra, and D. Manocha, “SS-SFDA:
+Self-supervised source-free domain adaptation for road segmen-
+tation in hazardous environments,” in Proceedings of the IEEE/CVF
+International Conference on Computer Vision, 2021, pp. 3049–3059.
+[198] D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell,
+“Tent: Fully test-time adaptation by entropy minimization,” arXiv
+preprint arXiv:2006.10726, 2020.
+[199] M. Bateson, H. Kervadec, J. Dolz, H. Lombaert, and I. Ben Ayed,
+“Source-relaxed domain adaptation for image segmentation,” in
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+18
+International Conference on Medical Image Computing and Computer-
+Assisted Intervention.
+Springer, 2020, pp. 490–499.
+[200] Y. Xu, J. Yang, H. Cao, K. Wu, M. Wu, and Z. Chen, “Source-
+free video domain adaptation by learning temporal consistency
+for action recognition,” in European Conference on Computer Vision.
+Springer, 2022, pp. 147–164.
+[201] Y. Huang, X. Yang, J. Zhang, and C. Xu, “Relative alignment net-
+work for source-free multimodal video domain adaptation,” in
+Proceedings of the 30th ACM International Conference on Multimedia,
+2022, pp. 1652–1660.
+[202] Y. Wang, J. Liang, and Z. Zhang, “Give me your trained model:
+Domain adaptive semantic segmentation without source data,”
+arXiv preprint arXiv:2106.11653, 2021.
+[203] J. Liang, D. Hu, J. Feng, and R. He, “UMAD: Universal model
+adaptation under domain and category shift,” arXiv preprint
+arXiv:2112.08553, 2021.
+[204] J. Dong, Z. Fang, A. Liu, G. Sun, and T. Liu, “Confident anchor-
+induced multi-source free domain adaptation,” Advances in Neu-
+ral Information Processing Systems, vol. 34, pp. 2848–2860, 2021.
+[205] Q. Tian, S. Peng, and T. Ma, “Source-free unsupervised domain
+adaptation with trusted pseudo samples,” ACM Transactions on
+Intelligent Systems and Technology, 2022.
+[206] R. M¨uller, S. Kornblith, and G. E. Hinton, “When does label
+smoothing help?” Advances in Neural Information Processing Sys-
+tems, vol. 32, 2019.
+[207] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna,
+“Rethinking the Inception architecture for computer vision,” in
+Proceedings of the IEEE Conference on Computer Vision and Pattern
+Recognition, 2016, pp. 2818–2826.
+[208] K. Li, J. Lu, H. Zuo, and G. Zhang, “Source-free multi-domain
+adaptation with generally auxiliary model training,” in 2022
+International Joint Conference on Neural Networks (IJCNN).
+IEEE,
+2022, pp. 1–8.
+[209] P. Chen and A. J. Ma, “Source-free temporal attentive domain
+adaptation for video action recognition,” in Proceedings of the 2022
+International Conference on Multimedia Retrieval, 2022, pp. 489–497.
+[210] Z. Xu, W. Wei, L. Zhang, and J. Nie, “Source-free domain
+adaptation for cross-scene hyperspectral image classification,”
+in IGARSS 2022-2022 IEEE International Geoscience and Remote
+Sensing Symposium.
+IEEE, 2022, pp. 3576–3579.
+[211] H. Yan, Y. Guo, and C. Yang, “Augmented self-labeling for
+source-free unsupervised domain adaptation,” in NeurIPS 2021
+Workshop on Distribution Shifts: Connecting Methods and Applica-
+tions, 2021.
+[212] C.-Y. Yang, Y.-J. Kuo, and C.-T. Hsu, “Source free domain adap-
+tation for semantic segmentation via distribution transfer and
+adaptive class-balanced self-training,” in 2022 IEEE International
+Conference on Multimedia and Expo (ICME).
+IEEE, 2022, pp. 1–6.
+[213] L. Xiong, M. Ye, D. Zhang, Y. Gan, and Y. Liu, “Source data-free
+domain adaptation for a faster R-CNN,” Pattern Recognition, vol.
+124, p. 108436, 2022.
+[214] N. Ma, J. Bu, L. Lu, J. Wen, Z. Zhang, S. Zhou, and X. Yan, “Semi-
+supervised hypothesis transfer for source-free domain adapta-
+tion,” arXiv preprint arXiv:2107.06735, 2021.
+[215] X. Guan, H. Sun, N. Liu, and H. Zhou, “Polycentric clustering
+and structural regularization for source-free unsupervised do-
+main adaptation,” arXiv preprint arXiv:2210.07463, 2022.
+[216] J. N. Kundu, A. R. Kulkarni, S. Bhambri, D. Mehta, S. A. Kulkarni,
+V. Jampani, and V. B. Radhakrishnan, “Balancing discriminability
+and transferability for source-free domain adaptation,” in Inter-
+national Conference on Machine Learning.
+PMLR, 2022, pp. 11 710–
+11 728.
+[217] C. Li, W. Chen, X. Luo, Y. He, and Y. Tan, “Adaptive pseudo
+labeling for source-free domain adaptation in medical image
+segmentation,” in ICASSP 2022-2022 IEEE International Conference
+on Acoustics, Speech and Signal Processing (ICASSP).
+IEEE, 2022,
+pp. 1091–1095.
+[218] Y. Kim, D. Cho, K. Han, P. Panda, and S. Hong, “Domain
+adaptation without source data,” IEEE Transactions on Artificial
+Intelligence, vol. 2, no. 6, pp. 508–518, 2021.
+[219] V. Prabhu, S. Khare, D. Kartik, and J. Hoffman, “S4T: Source-free
+domain adaptation for semantic segmentation via self-supervised
+selective self-training,” arXiv preprint arXiv:2107.10140, 2021.
+[220] M. Shen, Y. Bu, and G. Wornell, “On the benefits of selectivity
+in pseudo-labeling for unsupervised multi-source-free domain
+adaptation,” arXiv preprint arXiv:2202.00796, 2022.
+[221] T. Han, X. Gong, F. Feng, J. Zhang, Z. Sun, and Y. Zhang, “Privacy
+preserving mutli-source domain adaptaion for medical data,”
+IEEE Journal of Biomedical and Health Informatics, 2022.
+[222] X. Peng, Z. Huang, Y. Zhu, and K. Saenko, “Federated adversarial
+domain adaptation,” arXiv preprint arXiv:1911.02054, 2019.
+[223] H. Feng, Z. You, M. Chen, T. Zhang, M. Zhu, F. Wu, C. Wu,
+and W. Chen, “KD3A: Unsupervised multi-source decentralized
+domain adaptation via knowledge distillation.” in ICML, 2021,
+pp. 3274–3283.
+[224] Y. Kang, Y. He, J. Luo, T. Fan, Y. Liu, and Q. Yang, “Privacy-
+preserving federated adversarial domain adaptation over feature
+groups for interpretability,” IEEE Transactions on Big Data, 2022.
+[225] L. Song, C. Ma, G. Zhang, and Y. Zhang, “Privacy-preserving
+unsupervised domain adaptation in federated setting,” IEEE
+Access, vol. 8, pp. 143 233–143 240, 2020.
+[226] Z. Qin, L. Yang, F. Gao, Q. Hu, and C. Shen, “Uncertainty-aware
+aggregation for federated open set domain adaptation,” IEEE
+Transactions on Neural Networks and Learning Systems, 2022.
+[227] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman,
+V. Ivanov, C. Kiddon, J. Koneˇcn`y, S. Mazzocchi, B. McMahan
+et al., “Towards federated learning at scale: System design,”
+Proceedings of Machine Learning and Systems, vol. 1, pp. 374–388,
+2019.
+[228] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning:
+Challenges, methods, and future directions,” IEEE Signal Process-
+ing Magazine, vol. 37, no. 3, pp. 50–60, 2020.
+[229] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMa-
+han, S. Patel, D. Ramage, A. Segal, and K. Seth, “Practical
+secure aggregation for privacy-preserving machine learning,” in
+proceedings of the 2017 ACM SIGSAC Conference on Computer and
+Communications Security, 2017, pp. 1175–1191.
+[230] L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,”
+Advances in Neural Information Processing Systems, vol. 32, 2019.
+[231] A. Z. Tan, H. Yu, L. Cui, and Q. Yang, “Towards personalized
+federated learning,” IEEE Transactions on Neural Networks and
+Learning Systems, 2022.
+[232] C.-H. Yao, B. Gong, H. Qi, Y. Cui, Y. Zhu, and M.-H. Yang,
+“Federated multi-target domain adaptation,” in Proceedings of the
+IEEE/CVF Winter Conference on Applications of Computer Vision,
+2022, pp. 1424–1433.
+[233] D. Shenaj, E. Fan`ı, M. Toldo, D. Caldarola, A. Tavera, U. Michieli,
+M. Ciccone, P. Zanuttigh, and B. Caputo, “Learning across do-
+mains and devices: Style-driven source-free domain adaptation
+in clustered federated learning,” arXiv preprint arXiv:2210.02326,
+2022.
+[234] M. Yazdanpanah and P. Moradi, “Visual domain bridge: A
+source-free domain adaptation for cross-domain few-shot learn-
+ing,” in Proceedings of the IEEE/CVF Conference on Computer Vision
+and Pattern Recognition, 2022, pp. 2868–2877.
+[235] N. Karani, E. Erdil, K. Chaitanya, and E. Konukoglu, “Test-time
+adaptable neural networks for robust medical image segmenta-
+tion,” Medical Image Analysis, vol. 68, p. 101907, 2021.
+[236] M. Boudiaf, R. Mueller, I. Ben Ayed, and L.
+Bertinetto,
+“Parameter-free online test-time adaptation,” in Proceedings of the
+IEEE/CVF Conference on Computer Vision and Pattern Recognition,
+2022, pp. 8344–8353.
+[237] Y. Iwasawa and Y. Matsuo, “Test-time classifier adjustment mod-
+ule for model-agnostic domain generalization,” Advances in Neu-
+ral Information Processing Systems, vol. 34, pp. 2427–2440, 2021.
+[238] W. Ma, C. Chen, S. Zheng, J. Qin, H. Zhang, and Q. Dou, “Test-
+time adaptation with calibration of medical image classifica-
+tion nets for label distribution shift,” in International Conference
+on Medical Image Computing and Computer-Assisted Intervention.
+Springer, 2022, pp. 313–323.
+[239] F. You, J. Li, and Z. Zhao, “Test-time batch statistics calibration
+for covariate shift,” arXiv preprint arXiv:2110.04065, 2021.
+[240] Y. He, A. Carass, L. Zuo, B. E. Dewey, and J. L. Prince, “Au-
+toencoder based self-supervised test-time adaptation for medical
+image analysis,” Medical Image Analysis, vol. 72, p. 102136, 2021.
+[241] He, Yufan and Carass, Aaron and Zuo, Lianrui and Dewey,
+Blake E and Prince, Jerry L, “Self domain adapted network,” in
+International Conference on Medical Image Computing and Computer-
+Assisted Intervention.
+Springer, 2020, pp. 437–446.
+[242] H. Yang, C. Chen, M. Jiang, Q. Liu, J. Cao, P. A. Heng, and Q. Dou,
+“DLTTA: Dynamic learning rate for test-time adaptation on cross-
+domain medical images,” arXiv preprint arXiv:2205.13723, 2022.
+
+SOURCE-FREE UNSUPERVISED DOMAIN ADAPTATION: A SURVEY
+19
+[243] J. N. Kundu, N. Venkat, A. Revanur, R. V. Babu et al., “Towards
+inheritable models for open-set domain adaptation,” in Proceed-
+ings of the IEEE/CVF Conference on Computer Vision and Pattern
+Recognition, 2020, pp. 12 376–12 385.
+[244] Y. Zhao, Z. Zhong, Z. Luo, G. H. Lee, and N. Sebe, “Source-free
+open compound domain adaptation in semantic segmentation,”
+IEEE Transactions on Circuits and Systems for Video Technology,
+vol. 32, no. 10, pp. 7019–7032, 2022.
+[245] K. Saito, D. Kim, S. Sclaroff, and K. Saenko, “Universal domain
+adaptation through self supervision,” Advances in Neural Informa-
+tion Processing Systems, vol. 33, pp. 16 282–16 292, 2020.
+[246] Y. Luo, Z. Wang, Z. Chen, Z. Huang, and M. Baktashmotlagh,
+“Source-free progressive graph learning for open-set domain
+adaptation,” arXiv preprint arXiv:2202.06174, 2022.
+[247] R. Xu, Z. Chen, W. Zuo, J. Yan, and L. Lin, “Deep cocktail
+network: Multi-source unsupervised domain adaptation with
+category shift,” in Proceedings of the IEEE Conference on Computer
+Vision and Pattern Recognition, 2018, pp. 3964–3973.
+[248] E. Lekhtman, Y. Ziser, and R. Reichart, “DILBERT: Customized
+pre-training for domain adaptation with category shift, with
+an application to aspect extraction,” in Proceedings of the 2021
+Conference on Empirical Methods in Natural Language Processing,
+2021, pp. 219–230.
+[249] K. Saito and K. Saenko, “OVANet: One-vs-all network for uni-
+versal domain adaptation,” in Proceedings of the IEEE/CVF Inter-
+national Conference on Computer Vision, 2021, pp. 9000–9009.
+[250] Y. Feng, J. Chen, S. He, T. Pan, and Z. Zhou, “Globally localized
+multisource domain adaptation for cross-domain fault diagnosis
+with category shift,” IEEE Transactions on Neural Networks and
+Learning Systems, 2021.
+[251] T. Elsken, J. H. Metzen, and F. Hutter, “Neural architecture search:
+A survey,” The Journal of Machine Learning Research, vol. 20, no. 1,
+pp. 1997–2017, 2019.
+[252] M. Wistuba, A. Rawat, and T. Pedapati, “A survey on neural
+architecture search,” arXiv preprint arXiv:1905.01392, 2019.
+[253] Z. Lu, G. Sreekumar, E. Goodman, W. Banzhaf, K. Deb, and
+V. N. Boddeti, “Neural architecture transfer,” IEEE Transactions
+on Pattern Analysis and Machine Intelligence, vol. 43, no. 9, pp.
+2971–2989, 2021.
+[254] P. Ren, Y. Xiao, X. Chang, P.-Y. Huang, Z. Li, X. Chen, and
+X. Wang, “A comprehensive survey of neural architecture search:
+Challenges and solutions,” ACM Computing Surveys (CSUR),
+vol. 54, no. 4, pp. 1–34, 2021.
+[255] S. M. Ahmed, S. Lohit, K.-C. Peng, M. J. Jones, and A. K. Roy-
+Chowdhury, “Cross-modal knowledge transfer without task-
+relevant source data,” in European Conference on Computer Vision.
+Springer, 2022, pp. 111–127.
+[256] R. Kemker, M. McClure, A. Abitino, T. Hayes, and C. Kanan,
+“Measuring catastrophic forgetting in neural networks,” in Pro-
+ceedings of the AAAI Conference on Artificial Intelligence, vol. 32,
+no. 1, 2018.
+[257] J. Zhang, W. Li, C. Tang, P. Ogunbona et al., “Unsuper-
+vised domain expansion from multiple sources,” arXiv preprint
+arXiv:2005.12544, 2020.
+[258] M. Jing, X. Zhen, J. Li, and C. G. Snoek, “Variational model
+perturbation for source-free domain adaptation,” arXiv preprint
+arXiv:2210.10378, 2022.
+[259] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Des-
+jardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-
+Barwinska et al., “Overcoming catastrophic forgetting in neural
+networks,” Proceedings of the National Academy of Sciences, vol. 114,
+no. 13, pp. 3521–3526, 2017.
+[260] Z. Li and D. Hoiem, “Learning without forgetting,” IEEE Trans-
+actions on Pattern Analysis and Machine Intelligence, vol. 40, no. 12,
+pp. 2935–2947, 2017.
+[261] D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for
+continual learning,” Advances in Neural Information Processing
+Systems, vol. 30, 2017.
+[262] M. De Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia,
+A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual
+learning survey: Defying forgetting in classification tasks,” IEEE
+Transactions on Pattern Analysis and Machine Intelligence, vol. 44,
+no. 7, pp. 3366–3385, 2021.
+[263] S. Tang, P. Su, D. Chen, and W. Ouyang, “Gradient regularized
+contrastive learning for continual domain adaptation,” in Proceed-
+ings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 3,
+2021, pp. 2665–2673.
+[264] Q. Wang, O. Fink, L. Van Gool, and D. Dai, “Continual test-time
+domain adaptation,” in Proceedings of the IEEE/CVF Conference on
+Computer Vision and Pattern Recognition, 2022, pp. 7201–7211.
+[265] A. M. N. Taufique, C. S. Jahan, and A. Savakis, “ConDA:
+Continual unsupervised domain adaptation,” arXiv preprint
+arXiv:2103.11056, 2021.
+[266] B. Chidlovskii, S. Clinchant, and G. Csurka, “Domain adaptation
+in the absence of source domain data,” in Proceedings of the 22nd
+ACM SIGKDD International Conference on Knowledge Discovery and
+Data Mining, 2016, pp. 451–460.
+[267] A. R. Nelakurthi, R. Maciejewski, and J. He, “Source free domain
+adaptation using an off-the-shelf classifier,” in 2018 IEEE Interna-
+tional Conference on Big Data (Big Data).
+IEEE, 2018, pp. 140–145.
+[268] F. Wang, Z. Han, Z. Zhang, and Y. Yin, “Active source free domain
+adaptation,” arXiv preprint arXiv:2205.10711, 2022.
+[269] D. Kothandaraman, S. Shekhar, A. Sancheti, M.
+Ghuhan,
+T. Shukla, and D. Manocha, “DistillAdapt: Source-free active
+visual domain adaptation,” arXiv preprint arXiv:2205.12840, 2022.
+[270] B. Yang, A. J. Ma, and P. C. Yuen, “Revealing task-relevant model
+memorization for source-protected unsupervised domain adap-
+tation,” IEEE Transactions on Information Forensics and Security,
+vol. 17, pp. 716–731, 2022.
+[271] X. Wang, J. Zhuo, S. Cui, and S. Wang, “Learning invariant rep-
+resentation with consistency and diversity for semi-supervised
+source hypothesis transfer,” arXiv preprint arXiv:2107.03008, 2021.
+