A Source-Free Approach for Domain Adaptation via Multiview Image Transformation and Latent Space Consistency

Debopom Sutradhar, Md. Abdur Rahman, Mohaimenul Azam Khan Raiaan, Reem E. Mohamed, Sami Azam · 2026

Read A Source-Free Approach for Domain Adaptation via Multiview Image Transformation and Latent Space Consistency by Debopom Sutradhar, Md. Abdur Rahman, Mohaimenul Azam Khan Raiaan, Reem E. Mohamed, Sami Azam (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-20284.txt.

Introduction

Debopom Sutradhar and four colleagues present a source-free method for domain adaptation in image classification. Domain adaptation moves knowledge from a labeled source domain to a target domain where the data look different. Many methods need the source data at hand, or rely on adversarial training or heavy pseudo-labeling, all of which cost compute.

Their approach works without access to the source data. It uses multiview image transformation together with consistency in the latent space. Source-free methods suit real deployments, where privacy or storage rules often bar keeping the original training set.

Catalog record

CREATORDebopom Sutradhar, Md. Abdur Rahman, Mohaimenul Azam Khan Raiaan, Reem E. Mohamed, Sami Azam
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-20284

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Source-Free Approach for Domain Adaptation via Multiview Image Transformation and Latent Space Consistency Debopom Sutradhar, Md. Abdur Rahman, Mohaimenul Azam Khan Raiaan, Reem E. Mohamed, Sami Azam Abstract Domain adaptation (DA) addresses the challenge of transferring knowledge from a source domain to a target domain where image data distributions may differ. Existing DA methods often require access to source domain data, adversarial training, or complex pseudo-labeling techniques, which are computationally expensive. To address these challenges, this paper introduces a novel source-free domain adaptation method. It is the first approach to use multiview augmentation and latent space consistency techniques to learn domain-invariant features directly from the target domain. Our method eliminates the need for source-target alignment or pseudo-label refinement by learning transferable representations solely from the target domain by enforcing consistency between multiple augmented views in the latent space. Additionally, the method ensures consistency in the learned features by generating multiple augmented views of target domain data and minimizing the distance between their feature representations in the latent space. We also introduce a ConvNeXt-based encoder and design a loss function that combines classification and consistency objectives to drive effective adaptation directly from the target domain. The proposed model achieves an average classification accuracy of 90. 72%, 84%, and 97. 12% in Office-31, Office-Home and Office-Caltech datasets, respectively. Further evaluations confirm that our study improves existing methods by an average classification accuracy increment of +1.23%, +7.26%, and +1.77% on the respective datasets. I Introduction Image classification models often assume that training and testing data share the same underlying distribution. However, in real-world applications, variations in imaging conditions frequently result in domain shifts. These shifts degrade the performance of models trained in a single source domain when deployed on different target domains. Domain adaptation (DA) addresses this challenge by enabling models to generalize across domains with differing distributions. Recent advances in DA focus on aligning image feature distributions to mitigate domain shifts between source and target domains. A particularly promising method, source-free domain adaptation (SFDA), eliminates the need for source domain data during the training process. SFDA techniques typically rely on curriculum learning, pseudo-labeling, and entropy minimization [ 1 ] [ 2 ] [ 3 ] . However, pseudo-labeling can lead to error accumulation and confirmation bias, especially when early predictions are incorrect. Similarly, curriculum learning, while effective for reducing label noise, may neglect more challenging examples, limiting the model’s ability to adapt fully. Additionally, the absence of source supervision makes SFDA methods prone to catastrophic forgetting, where the model loses knowledge learned from the source domain, weakening its ability to retain transferable features. To address these challenges, curriculum learning progressively incorporates high-confidence pseudo-labels to reduce label noise, ensuring that the model adapts more effectively. Similarly, dynamic pseudo-labeling strategies and entropy-based methods are used to minimize uncertainty in generated labels to further improve adaptation performance [ 1 ] [ 4 ] . Additionally, memory-based methods also store class prototypes learned during source training, align target samples with these prototypes, and enable adaptation without direct reliance on source-target comparisons. Additionally, multi-modal and fusion strategies have been developed to handle complex environments involving diverse data modalities. Generative models, such as GANs, are used for pretraining by synthesizing realistic representations of complex data distributions. For example, GAN-based frameworks generate synthetic samples that preserve intermodal correlations, which are later fine-tuned for downstream fusion tasks [ 5 ] . These tasks often involve learning unified feature representations that can seamlessly integrate information from disparate sources. In contrast, energy-based models (EBMs) capture latent dependencies between multimodal data sources in some works [ 6 ] . These approaches aim to handle multisource and multimodal data and improve feature extraction by learning shared representations across various data types. Advanced methods often integrate attention mechanisms into these frameworks to enable the model to dynamically weigh contributions from each modality based on their relevance to the task. DA methods have also been extended to address scenarios with highly imbalanced data distributions. Domain-specific loss functions penalize the misclassification of underrepresented classes more heavily to ensure that the model focuses more on these difficult cases [ 7 ] . Revisiting mechanisms adjust sample weights according to class distribution to further mitigate the impact of imbalanced data. Furthermore, regularization techniques such as entropy minimization or adversarial training are used to prevent overfitting and ensure that the model generalizes better to diverse and skewed datasets [ 8 ] . To overcome existing challenges, this study introduces a novel Source-free Domain Adaptation approach. Unlike single-view transformation strategies, our framework employs a multiview augmentation strategy to enable the model to observe diverse views of the same sample and learn consistent representations across them to identify domain-invariant features. To the best of our knowledge, we are the first to integrate multiview augmentation, latent space consistency, and source-free operation into a unified framework that learns domain-invariant features directly from the target domain without relying on source supervision. Our method further eliminates reliance on adversarial setups and pseudo-labeling as well. The major contributions of our study are as follows. • We propose a novel multiview augmentation strategy that generates multiple augmented views for each image in the target domain. Random transformations are applied to simulate different imaging conditions to help the model learn domain-invariant features without requiring source data. • We introduce a novel latent space consistency mechanism in which latent feature representations of augmented views are aligned in the latent space. The model learns consistent features and addresses domain shifts by minimizing the distance between augmented view representations. • An objective function is incorporated that combines classification and consistency losses to retain discriminative power for target domain samples while ensuring feature consistency across augmented views. • A ConvNeXt-based encoder is used for its feature extraction capabilities via depthwise convolutions to map the input data to a latent space for domain adaptation without adversarial training. The encoder is customized to reduce dimensionality and focus on essential features for domain adaptation. The remainder of this paper is organized as follows. Section II reviews recent related work in domain adaptation. Section III presents our proposed methodology, including problem formulation and model architecture. Section IV provides experimental setup and training details. Section V details the analysis of the results, performance comparisons, and ablation studies. Section VI provides a discussion of the study, including limitations and potential future improvements. Finally, Section VII concludes the paper by summarizing our key findings. II Related Works This section provides a review of recent domain adaptation methods, which we have grouped into four primary categories: pseudolabeling-based techniques, representation learning methods, adversarial and contrastive learning approaches, and multiview augmentation techniques. Pseudo-labeling is an effective technique in the adaptation of the source-free domain, where the provisional labels assigned to the target samples guide the refinement of the features and the alignment of the domain using the predictions of the model to iteratively adapt to the target distribution without requiring access to source data. Recent advances [ 9 , 10 , 11 , 12 , 13 ] have introduced innovative refinements to enhance the effectiveness of pseudo-labeling. For example, Liang et al. [ 9 ] introduced SHOT, which focuses on aligning target features through information maximization and self-supervised learning. Yang et al.

Read the complete text