Source-Guided Target Feature Reconstruction for Cross-Domain Classification and Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised cross-domain learning methods often overlook crucial source domain knowledge.
• Existing techniques rely on self-augmentation for consistency, limiting performance in domain adaptation.

Purpose of the Study:

• To propose a Source-guided Target Feature Reconstruction (STFR) module for enhancing cross-domain visual tasks.
• To reduce domain bias and improve target representation by integrating source knowledge.

Main Methods:

• The STFR module reconstructs target features using source visual words, creating a bridge between domains.
• Source visual words are dynamically selected and updated based on source feature distribution.
• Consistency learning is applied between reconstructed and original target features for domain alignment, theoretically linked to optimal transport.

Main Results:

• Achieved 91.0% accuracy on Office-31, 73.9% on Office-Home, and 87.4% on VisDA-2017 for cross-domain image classification.
• Obtained 44.50% mAP on Cityscapes → Foggy Cityscapes and 78.10% AP for car on Cityscapes → KITTI for cross-domain object detection.
• Demonstrated significant improvements across multiple benchmarks for both classification and detection tasks.

Conclusions:

• The proposed STFR module effectively enhances cross-domain visual tasks by integrating source domain knowledge.
• STFR significantly reduces domain bias, leading to superior performance in unsupervised domain adaptation.
• The method proves effective across diverse cross-domain image classification and object detection benchmarks.