Diverse Target and Contribution Scheduling for Domain Generalization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generalization under distribution shifts is a major challenge in computer vision.
• Current domain generalization (DG) methods using one-hot labels can cause gradient conflicts and fail to capture intrinsic class characteristics.
• Existing DG approaches often neglect the distinct contributions of source domains, leading to imbalanced learning.

Purpose of the Study:

• To address gradient conflicts and uneven source domain learning in DG.
• To propose a novel paradigm for DG called Diverse Target and Contribution Scheduling (DTCS).
• To analyze the relationship between distribution shifts and gradient conflicts in DG.

Main Methods:

• Proposed Diverse Target and Contribution Scheduling (DTCS) paradigm for DG.
• Introduced Diverse Target Supervision (DTS) using distinct soft labels to mitigate gradient conflicts.
• Implemented Diverse Contribution Balance (DCB) to dynamically balance source domain contributions.

Main Results:

• DTCS effectively addresses limitations of one-hot labels and equal source domain contributions in DG.
• Experiments on four benchmark datasets demonstrate competitive performance against state-of-the-art methods.
• The proposed method shows effectiveness and advantages in domain generalization tasks.

Conclusions:

• DTCS offers a novel perspective on DG by considering empirical source domain risk.
• The method successfully mitigates gradient conflicts and improves intra-class variation.
• DTCS represents a significant advancement in domain generalization research.