Class-Customized Domain Adaptation: Unlock Each Customer-Specific Class With Single Annotation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• General-purpose AI models often perform poorly in specialized domains, leading to inefficiencies and privacy concerns.
• Existing domain adaptation methods struggle with scenarios requiring customization for unique, customer-specific classes.
• Low-cost annotation remains a significant hurdle for effective model customization.

Purpose of the Study:

• To propose a novel Class-Customized Domain Adaptation (CCDA) method for efficient model customization.
• To address the challenge of adapting models to specialized domains with customer-specific classes using minimal annotations.
• To improve model performance and reduce resource wastage in domain-specific AI applications.

Main Methods:

• Developed a Class-Customized Domain Adaptation (CCDA) framework utilizing a classic adaptation training approach.
• Introduced a partial-feature alignment strategy to ensure accurate knowledge transfer from shared and private classes.
• Implemented soft-balanced sampling to mitigate the long-tail distribution problem in customer-specific class data.

Main Results:

• CCDA demonstrated consistent and excellent performance across 48 simulated domain adaptation tasks.
• The method proved effective in two real-world customization scenarios.
• Analytical experiments validated the significant contributions of the partial-feature alignment and soft-balanced sampling techniques.

Conclusions:

• The proposed CCDA method offers an effective solution for low-cost model customization in specialized domains, even with customer-specific classes.
• CCDA enhances model performance and resource efficiency by enabling precise adaptation with minimal annotations.
• This research advances domain adaptation techniques, particularly for scenarios with unique class requirements.