Smooth-Guided Implicit Data Augmentation for Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization (DG) aims to train models on source domains for optimal performance on unseen target domains.
• Current DG methods often require auxiliary networks or high computational costs to integrate diverse source domains.
• Improving model generalization ability remains a key challenge in machine learning.

Purpose of the Study:

• To propose a novel method, Smooth-Guided Implicit Data Augmentation (SGIDA), for enhancing domain generalization.
• To capture source domain diversity efficiently in the feature space.
• To improve model generalization capacity without relying solely on deep features.

Main Methods:

• SGIDA operates in the feature space to capture source domain diversity.
• Incorporates a distance metric learning (DML) loss function to amplify generalization capacity.
• Utilizes logits from cross-entropy (CE) losses with infinite augmentations, avoiding reliance on deep features.
• Introduces a sampling-based method called 'smooth' to obtain semantic directions from interclass relations for increased source domain diversity.

Main Results:

• Theoretical analysis demonstrates the effectiveness of logits in estimating distances on original features for DG.
• Extensive experiments show significant improvements over state-of-the-art methods on DG, object detection, and remote sensing datasets.
• The proposed approach achieves superior performance across various backbone networks.

Conclusions:

• SGIDA offers an effective and computationally efficient approach to domain generalization.
• The use of logits and smooth sampling enhances model adaptability to unseen target domains.
• The method demonstrates broad applicability and significant performance gains in computer vision tasks.