Foreground-aware Universe Graph Matching for Domain Adaptive Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Domain Adaptive Object Detection (DAOD) aims to improve model generalization across different data domains.
• Current DAOD methods often align cross-domain distributions using graph matching but struggle with precise semantic alignment.
• Challenges include domain-biased graph modeling and unreliable matching with background nodes.

Purpose of the Study:

• To propose a novel framework, Foreground-aware Universe Graph Matching (FUGM), to address limitations in current DAOD.
• To enhance semantic alignment by focusing on foreground object instances.
• To improve the generalization ability of object detection models in cross-domain scenarios.

Main Methods:

• Constructing a virtual universe graph for category-wise semantic knowledge modeling.
• Refining node representations using Collaborative Graph Reasoning (CGR) with self-loop calibration.
• Developing Universe Graph Matching (UGM) to encourage instance-anchor node matching and reduce spurious pairs.

Main Results:

• The FUGM framework enables end-to-end learning for enhanced instance node pairwise affinity.
• Gathering foreground nodes of corresponding categories improves matching accuracy.
• Extensive experiments demonstrate significant performance improvements over existing DAOD methods.

Conclusions:

• FUGM effectively overcomes the limitations of previous graph-matching approaches in DAOD.
• The proposed method achieves superior semantic alignment and cross-domain generalization.
• FUGM represents a significant advancement in the field of domain adaptive object detection.