OriMamba: Remote sensing oriented object detection with state space models

Oriented object detection plays a pivotal role in remote sensing image monitoring and understanding, such as traffic monitoring and urban construction. However, it poses notable difficulties stemming from the varied angles, scales, and shapes of objects, as well as complex backgrounds. In this study, we introduce OriMamba, an innovative framework for detecting oriented objects in RSIs, leveraging a novel state-space model architecture. To efficiently address the challenges posed by objects of differing scales and complex backgrounds, we design a hybrid Mamba pyramid network that integrates orientation and multi-scale features by combining global and local information. Furthermore, we propose a dynamic double head structure, which separates the spatial prediction and classification tasks to enhance the accuracy of object orientation estimation. Experimental results show that our method achieves 78.23 mAP on DOTA-1.0, 70.90 mAP on DOTA-1.5, 90.70 mAP on HRSC2016, and 64.53 mAP on DIOR-R, surpassing existing state-of-the-art methods by 0.5, 2.88, 0.1, and 2.38 mAP, respectively.