Improving Remote Sensing Object Detection Based on Relation DETR

Relation-DETR is a model that integrates positional relationships into the DETR framework, iteratively refining bounding box predictions. Its advantages include mitigating scale bias, enhancing performance and convergence speed, and demonstrating strong transferability. The model optimizes DETR's self-attention mechanism, enabling it to globally leverage contextual information and subsequently perform object detection based on these relationships. However, the computation of attention weights necessitates additional computational resources, thereby increasing processing time and hardware consumption when handling data. To address this issue, this study proposes an improved dynamic convolution approach. By adopting a simplified computational method, it effectively incorporates Relation-DETR, enhancing the model's capabilities, facilitating training, and accelerating training speed. We evaluated its performance on the NWPU VHR-10 dataset during training and achieved an average precision (AP) of 73.1% after only 12 epochs, which is comparable to existing state-of-the-art methods. This approach holds promise for providing additional techniques and is convenient for future research.