Scholarlink

<p>Source-free Domain Adaptation aims to adapt a pre-trained source model to an unlabeled target domain while circumventing access to well-labeled source data. To compensate for the absence of source data, most existing approaches employ prototype-based pseudo-labeling strategies to facilitate self-training model adaptation. Nevertheless, these methods commonly rely on instance-level predictions for direct monocentric prototype construction, leading to category bias and noisy labels. This is primarily due to the inherent visual domain gaps that often differ across categories. Besides, the monocentric prototype design is ineffective and may introduce negative transfer for those ambiguous data. To tackle these challenges, we propose a general class-<b>B</b>alanced <b>M</b>ulticentric <b>D</b>ynamic (BMD) prototype strategy. Specifically, we first introduce a global inter-class balanced sampling strategy for each target category to mitigate category bias. Subsequently, we design an intra-class multicentric clustering strategy to generate robust and representative prototypes. In contrast to existing approaches that only update pseudo-labels at fixed intervals, e.g., one epoch, we employ a dynamic pseudo-labeling strategy that incorporates network update information throughout the model adaptation. We refer to the vanilla implementation of these three sub-strategies as BMD-v1. Furthermore, we promote the BMD-v1 to BMD-v2 by incorporating a consistency-guided reweighting strategy to improve inter-class balanced sampling, and leveraging the silhouettes metric to realize adaptive intra-class multicentric clustering. Extensive experiments conducted on both 2D images and 3D point cloud recognition demonstrate that our proposed BMD strategy significantly improves existing representative methods. Remarkably, BMD-v2 improves NRC from 52.6 to 59.2% in accuracy on the PointDA-10 benchmark. The code will be available at <a href="https://github.com/ispc-lab/BMD">https://github.com/ispc-lab/BMD</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

General Class-Balanced Multicentric Dynamic Prototype Pseudo-Labeling for Source-Free Domain Adaptation

  • 期刊论文

  • 发表日期：

    2025年1月5日

查看全文
  • Sanqing Qu,
  • Guang Chen,
  • Jing Zhang,
  • Zhijun Li,
  • Wei He,
  • Dacheng Tao

摘要

Source-free Domain Adaptation aims to adapt a pre-trained source model to an unlabeled target domain while circumventing access to well-labeled source data. To compensate for the absence of source data, most existing approaches employ prototype-based pseudo-labeling strategies to facilitate self-training model adaptation. Nevertheless, these methods commonly rely on instance-level predictions for direct monocentric prototype construction, leading to category bias and noisy labels. This is primarily due to the inherent visual domain gaps that often differ across categories. Besides, the monocentric prototype design is ineffective and may introduce negative transfer for those ambiguous data. To tackle these challenges, we propose a general class-Balanced Multicentric Dynamic (BMD) prototype strategy. Specifically, we first introduce a global inter-class balanced sampling strategy for each target category to mitigate category bias. Subsequently, we design an intra-class multicentric clustering strategy to generate robust and representative prototypes. In contrast to existing approaches that only update pseudo-labels at fixed intervals, e.g., one epoch, we employ a dynamic pseudo-labeling strategy that incorporates network update information throughout the model adaptation. We refer to the vanilla implementation of these three sub-strategies as BMD-v1. Furthermore, we promote the BMD-v1 to BMD-v2 by incorporating a consistency-guided reweighting strategy to improve inter-class balanced sampling, and leveraging the silhouettes metric to realize adaptive intra-class multicentric clustering. Extensive experiments conducted on both 2D images and 3D point cloud recognition demonstrate that our proposed BMD strategy significantly improves existing representative methods. Remarkably, BMD-v2 improves NRC from 52.6 to 59.2% in accuracy on the PointDA-10 benchmark. The code will be available at https://github.com/ispc-lab/BMD.