Dual Instance-Consistent Network for Cross-Domain Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Cross-domain object detection transfers knowledge from labeled source data to unlabeled target data.
• Existing methods often use unified embedding models, causing target domain feature distribution bias.
• This bias arises because the embedding model is primarily influenced by the source domain.

Purpose of the Study:

• To propose a novel Dual Instance-Consistent Network (DICN) for cross-domain object detection.
• To reduce representation bias in the target domain by employing two independent networks.
• To achieve mutual consistency between source-specific and target-specific descriptions.

Main Methods:

• Introduced the Dual Instance-Consistent Network (DICN) with a Dual Instance-Consistent Module.
• Utilized Primary and Auxiliary Networks to extract source-specific and target-specific information, respectively.
• Implemented instance mutual consistency through feature consistency and detection consistency for domain alignment.

Main Results:

• Achieved 44.10% mAP for Cityscapes → Foggy Cityscapes.
• Obtained 76.50% AP on car for Cityscapes → KITTI.
• Demonstrated significant improvements on person detection benchmarks, including COCOPersons → Caltech and others.

Conclusions:

• The proposed DICN effectively reduces target domain representation bias in cross-domain object detection.
• Instance mutual consistency between feature and detection levels ensures robust domain alignment.
• DICN shows superior performance across multiple challenging cross-domain object detection benchmarks.