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

To address the challenges of diverse object morphologies and significant inter-domain distribution differences in cross-domain object detection, this paper introduces the integration of Deformable Large Kernel Attention (DLKA) into the RT-DETR detection model, enhancing performance through the innovative combination of large convolution kernels and deformable convolutions. Specifically, large convolution kernels employ depth-wise separable convolutions and dilation techniques to expand the receptive field for capturing rich contextual information while reducing computational costs, effectively mimicking the global feature modeling capability of self-attention mechanisms. Deformable convolutions dynamically learn sampling offsets to adaptively adjust the sampling positions of convolution kernels, enhancing the model's adaptability to irregular object shapes and complex layouts in cross-domain scenarios. Experimental results demonstrate that the model incorporating DLKA achieves improvements of approximately 6.1%, 5.5%, and 5.6% in mAP50, Recall, and Precision metrics, respectively, compared to the baseline RT-DETR-resnet18. Notably, it exhibits more substantial performance advantages in late-stage training. This mechanism fundamentally enhances the model's robustness to object morphology and distribution differences across domains, significantly improving detection accuracy and recall in cross-domain scenarios. It provides an effective solution to address domain discrepancies in cross-domain object detection.