Domain Invariant and Class Discriminative Feature Learning for Visual Domain Adaptation

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Domain adaptation addresses differing data distributions between source and target domains.
• Existing methods focus on domain-invariant features but often neglect class discriminability.

Purpose of the Study:

• To propose a novel feature learning method for domain adaptation called DICD.
• To develop representations that are both domain-invariant and class-discriminative.

Main Methods:

• DICD learns a latent feature space by matching marginal and class-conditional distributions.
• It simultaneously maximizes inter-class dispersion and minimizes intra-class scatter.
• The method is optimized via a generalized eigen-decomposition problem.

Main Results:

• Class discriminative properties significantly reduce cross-domain distribution inconsistency.
• DICD demonstrates superior performance over existing methods in visual cross-domain classification tasks.
• Integrating domain invariance and class discriminativeness enhances classification accuracy.

Conclusions:

• DICD effectively learns domain-invariant and class-discriminative representations for domain adaptation.
• The proposed approach offers a significant improvement in cross-domain classification tasks.
• The method provides a unified optimization framework for feature learning in domain adaptation.