SIGMA++: Improved Semantic-Complete Graph Matching for Domain Adaptive Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain Adaptive Object Detection (DAOD) aims to generalize object detectors to new domains without labels.
• Current prototype-based methods struggle with class variance and domain-mismatched classes.
• Existing approaches often fail to capture structural dependencies and optimize adaptation for all classes.

Purpose of the Study:

• To propose an improved framework, SIGMA++, for Domain Adaptive Object Detection (DAOD).
• To address limitations of prototype-based methods in capturing class variance and handling domain-mismatched classes.
• To enhance cross-domain adaptation by completing semantics and employing hypergraph matching.

Main Methods:

• Introduced a Hypergraphical Semantic Completion (HSC) module to generate nodes for mismatched classes.
• Modeled class conditional distribution with high-order dependencies using cross-image hypergraphs.
• Reformulated domain adaptation as a hypergraph matching problem solved by a Bipartite Hypergraph Matching (BHM) module.
• Utilized graph nodes for semantic-aware affinity and edges for structural constraints in a structure-aware matching loss.

Main Results:

• SIGMA++ demonstrates state-of-the-art performance on nine benchmarks for DAOD.
• Achieved significant improvements in both AP50 and adaptation gains.
• Validated the generalization capability of SIGMA++ across various object detectors.

Conclusions:

• SIGMA++ effectively addresses limitations in prototype-based DAOD by incorporating hypergraph matching.
• The proposed HSC and BHM modules enable fine-grained adaptation and improved performance.
• The framework shows strong generalization and achieves leading results in domain adaptive object detection.