Academic Background and Problem Statement

In recent years, deep learning models (Deep Neural Networks, DNNs) have achieved remarkable success in computer vision tasks. However, the training of these models relies heavily on large amounts of annotated data. When models are applied to new, unlabeled target domains, their generalization ability often suffers due to domain shifts. To address this issue, Domain Adaptation (DA) techniques have emerged. The goal of DA is to leverage knowledge from a source domain to improve model performance in a target domain, especially when the target domain lacks labeled data.

However, traditional DA methods typically require access to the source domain’s raw data, which is often impractical in many real-world applications due to concerns about data privacy, security, and transmission efficiency. To tackle this problem, Source-Free Domain Adaptation (SFDA) has become an emerging research direction. SFDA aims to adapt a pre-trained source model to a target domain using only unlabeled target data, without accessing the source domain’s data.

Existing SFDA methods can be broadly divided into two categories: generative-based methods and self-training-based methods. Generative-based methods use Generative Adversarial Networks (GANs) or diffusion models to generate source-style images, while self-training-based methods assign pseudo-labels to target data using the source model. However, existing self-training methods often rely on monocentric prototypes to generate pseudo-labels, which can lead to category bias and noisy labels, especially when there are significant visual domain gaps between different categories.

To address these issues, this paper proposes a General Class-Balanced Multicentric Dynamic Prototype Pseudo-Labeling Strategy (BMD). This strategy significantly improves the performance of existing methods by introducing global class-balanced sampling, intra-class multicentric clustering, and dynamic pseudo-label generation.

Paper Source and Author Information

This paper is co-authored by Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. The authors are affiliated with Tongji University, Wuhan University, University of Science and Technology Beijing, and Nanyang Technological University. The paper was published in the International Journal of Computer Vision and officially released in 2025.

Research Methodology and Process

The proposed BMD strategy consists of three sub-strategies: Global Inter-Class Balanced Sampling, Intra-Class Multicentric Clustering, and Dynamic Pseudo-Labeling. Below is a detailed description of each sub-strategy:

1. Global Inter-Class Balanced Sampling

To prevent easy-transfer classes from dominating the prototype generation process, this paper introduces a global inter-class balanced sampling strategy. Specifically, for each target category, the most likely samples are selected from the target data, and these samples are averaged to construct a class-balanced feature prototype. Unlike existing methods, this strategy selects the most representative samples from a global perspective rather than relying on local instance-level predictions.

2. Intra-Class Multicentric Clustering

To reduce the impact of noisy labels, this paper proposes an intra-class multicentric clustering strategy. Unlike the existing monocentric prototype design, this strategy generates multiple feature prototypes for each category, providing more flexible and precise decision boundaries for pseudo-label assignment. Specifically, the classical K-means algorithm is used to cluster samples within each category, generating multiple feature prototypes.

3. Dynamic Pseudo-Labeling

Existing methods typically update pseudo-labels at fixed intervals (e.g., per epoch), which limits the effective utilization of network updates during training. To address this issue, this paper proposes a dynamic pseudo-labeling strategy based on Exponential Moving Average (EMA). This strategy updates pseudo-labels at the feature level, thereby improving model performance without significantly increasing computational costs.

Experimental Results and Contributions

Extensive experiments were conducted on multiple 2D image and 3D point cloud recognition datasets to validate the effectiveness and generality of the BMD strategy. The results show that BMD significantly improves the performance of existing methods. For example, on the PointDA-10 dataset, BMD-v2 increased the accuracy of the NRC method from 52.6% to 59.2%.

The main contributions of this paper can be summarized as follows: 1. A general Class-Balanced Multicentric Dynamic Prototype Strategy (BMD) is proposed, which is model-agnostic and can be applied to existing self-training-based SFDA methods. 2. A simple yet effective global inter-class balanced sampling strategy is introduced to prevent easy-transfer classes from dominating the prototype generation process. 3. An intra-class multicentric clustering strategy is proposed to generate multiple feature prototypes for each category, providing more precise decision boundaries for pseudo-label assignment. 4. A dynamic pseudo-labeling strategy is employed to fully utilize the model’s update information during training, further enhancing model performance.

Research Highlights and Significance

The highlights of this research include: 1. Class-Balanced Sampling: The global inter-class balanced sampling strategy effectively avoids category bias, improving the model’s generalization ability. 2. Multicentric Clustering: The intra-class multicentric clustering strategy generates more flexible and precise decision boundaries for pseudo-label assignment, reducing the impact of noisy labels. 3. Dynamic Pseudo-Labeling: The dynamic pseudo-labeling strategy fully leverages the model’s update information during training, further enhancing model performance.

This research not only holds significant scientific value but also provides effective solutions for domain adaptation problems in practical applications. Especially in the context of increasing concerns about data privacy and security, SFDA technology has broad application prospects.

Conclusion and Future Work

This paper proposes a general Class-Balanced Multicentric Dynamic Prototype Strategy (BMD) for source-free domain adaptation tasks. By introducing global inter-class balanced sampling, intra-class multicentric clustering, and dynamic pseudo-labeling, BMD significantly improves the performance of existing methods. Future work will explore applying this strategy to other visual tasks, such as semantic segmentation and object detection.