Universal Synchronization Loss Optimization in DETR-Based Oriented and Rotated Object Detection

The network models based on Detection Transformer (DETR) have gained popularity in object detection tasks. However, their performance is hindered by the lack of explicit correlation between classification and localization losses, resulting in suboptimal optimization and convergence. To address this issue, we propose universal synchronization loss. It can associate different positioning losses to improve detection accuracy. Specifically, we introduce L1 distance on the oriented object detection dataset and control the GIOU loss through a function. In the rotated dataset, we introduce the L1 distance to control the Kullback-Leibler divergence loss. In addition, we propose a method to dynamically adjust the classification weights to further associate the relationship between classification loss and localization loss. Experimental results on the COCO and DOTA v1.0/v1.5/v2.0 dataset demonstrate that our proposed universal synchronization loss brings excellent performance, surpassing previous DETR models and other similar models. This work underscores the importance of synchronizing losses in improving the overall optimization and detection accuracy of DETR-based models. Our code is available at https://github.com/yuqingchen1/SL-DETR.