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

Abstract

The excellence of intelligent detection models has been widely recognized, but in terms of cross-domain scenes, they still face performance degradation and low accuracy. A multi-supervised Tri-Flow-YOLO model is proposed to improve the accuracy of objects with various scales under cross-domain conditions. Based on the full-supervised traditional detection branch of YOLOv5, another two mutually supporting task branches are designed intently. In brief, we add unsupervised adversarial classification training flow to the backend, to realize the feature alignment requirements and improve the cross-domain performance stability of the model. Meanwhile, a weakly-supervised object counting flow is proposed to improve the model's attention to all the objects and the detection ability is efficiently enforced. In addition, I-Mosaic and iCIOU are designed especially for small hard objects, enriching the positive samples during the training process. With the auxiliary of both improved strategies, the imbalance of positive and negative samples in the anchor-based model is relieved accordingly. The experimental results show that the improved Tri-Flow-YOLO model achieves 56.0 mAP in the Cityscapes→Foggy-Cityscapes task, and 49.8 mAP in the VOC→Clipart task.

Keywords

• Cross-domain object detection
• Multi-task
• Mosaic
• Unsupervised domain adaptation