DiffRotDet: Diffusion-Based Rotated Object Detection Network

Since objects in remote sensing images captured from the aerial perspective exhibit arbitrary orientations, traditional horizontal detection struggles to represent rotated features. This necessitates the development of rotated detection frameworks in remote sensing. This paper proposes a novel Diffusion-based framework tailored for Rotated object Detection in remote sensing contexts, named DiffRotDet. The first advantage is that DiffRotDet formulates the detection task as a denoising diffusion process applied to rotated bounding boxes. By parameterizing object orientation—a defining attribute of rotated bounding boxes—and employing the RoI Align Rotated, our approach successfully extends the diffusion-based framework from horizontal to rotated object detection. This formulation avoids the need for heuristic anchor size selection, enabling robust adaptation to targets exhibiting diverse orientations and aspect ratios. Furthermore, we leverage the KLD (Kullback-Leibler Divergence) to construct the loss function to address the angle periodicity problem brought by the target representation method. This design significantly enhances the model's capability in detecting arbitrarily oriented objects, constituting the second key advantage of DiffRotDet. Experimental results on the public DIOR-R dataset show that DiffRotDet achieves 65.9 % mAP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf>, outperforming the state-of-the-art methods.