Optimization of Cross-Domain Detection Capabilities Based on RT-DETR

Object detectors often suffer a significant performance decline when faced with domain shifts between the source domain (collected data) and the target domain (actual application data). This is due to significant visual differences between images across domains, such as variations in object scale, texture, and content style. To improve cross-domain detection performance, this paper proposes integrating two modules: AssemFormer (an assembly-based convolutional vision transformer) and SEAM (Separated and Enhanced Attention Module) into the RT-DETR detector. AssemFormer combines the local feature extraction capabilities of convolutional neural networks with the global context modeling power of Transformers, addressing the limitations of traditional convolutional neural networks in capturing long-range dependencies and local details. SEAM improves feature responses in unobstructed regions while compensating for information loss in occluded areas, thereby enhancing detection capabilities for obscured objects. It also addresses the lack of inductive bias and weak local detail capture in pure Transformers. Together, these modules mitigate the adverse effects of domain differences between synthetic and real images, optimizing performance for cross-domain object detection. In the Sim10k-Cityscapes crossdomain detection task, the mAP improved by 5.8%, and in the Cityscapes-FoggyCityscapes task, it increased by 5.7%.