Cross-Domain Object Detection with Hierarchical Multi-Scale Domain Adaptive YOLO

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain shift degrades object detection performance.
• Existing domain adaptive object detection (DAOD) methods often use inefficient two-stage detectors.
• There is a need for efficient DAOD methods suitable for real-world applications.

Purpose of the Study:

• To propose a novel Hierarchical Multi-scale Domain Adaptive (HMDA) method integrated with a one-stage YOLO framework.
• To improve object detection performance on target domains despite domain shift.
• To achieve real-time detection efficiency.

Main Methods:

• Developed HMDA-YOLO, a one-stage domain adaptive object detection method.
• Implemented hierarchical backbone adaptation to align feature distributions across different network depths.
• Incorporated multi-scale head adaptation to leverage feature map information for improved detection.

Main Results:

• HMDA-YOLO demonstrates competitive performance across various cross-domain object detection scenarios.
• The method effectively reduces distribution discrepancies between domains.
• Achieved real-time detection efficiency without compromising accuracy.

Conclusions:

• HMDA-YOLO offers an effective solution for domain adaptive object detection.
• The proposed hierarchical and multi-scale adaptations improve generalization and discrimination capabilities.
• HMDA-YOLO provides a practical and efficient approach for real-world object detection challenges.