Tri-Flow-YOLO: Counter helps to improve cross-domain object detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Intelligent detection models exhibit performance degradation and low accuracy in cross-domain scenarios.
• Existing models face challenges in maintaining accuracy across diverse visual environments.

Purpose of the Study:

• To propose a multi-supervised Tri-Flow-YOLO model for improved object detection accuracy in cross-domain scenes.
• To enhance the stability and robustness of detection models when applied to unfamiliar visual domains.

Main Methods:

• The Tri-Flow-YOLO model integrates a full-supervised YOLOv5 branch with unsupervised adversarial classification and weakly-supervised object counting branches.
• Feature alignment is achieved through unsupervised adversarial classification for improved cross-domain performance stability.
• Object counting flow enhances model attention and detection capabilities for all objects.
• I-Mosaic and iCIOU strategies are introduced to enrich positive samples, particularly for small objects, and address sample imbalance.

Main Results:

• The enhanced Tri-Flow-YOLO model achieved 56.0 mAP on the Cityscapes→Foggy-Cityscapes cross-domain task.
• The model attained 49.8 mAP on the VOC→Clipart cross-domain task, demonstrating significant improvements.
• The proposed methods effectively improved accuracy for objects of various scales under cross-domain conditions.

Conclusions:

• The Tri-Flow-YOLO model significantly improves object detection accuracy and stability in cross-domain scenarios.
• Multi-supervised learning, incorporating adversarial and counting flows, is effective for enhancing model generalization.
• The specialized strategies for small objects and sample imbalance contribute to the overall performance gains.