ODANet: Lightweight Real-Time Oriented Object Detection via Oriented Depthwise Attention in Remote Sensing Imagery

Real-time oriented object detection in highresolution remote sensing imagery is critical for Earth observation applications, yet remains challenging due to arbitrary object orientations, complex scene layouts, and strict computational constraints on edge platforms. This paper presents ODANet, a lightweight oriented object detection framework that enhances orientation robustness at the feature representation level within a unified detection pipeline. Specifically, a rotation-invaiant attention mechanism is introduced to mitigate feature sensitivity to object orientation, while a lightweight dual-path enhancement strategy strengthens spatial context modeling and channel-wise representation capability. The overall design enables effective orientation-aware feature learning while introducing only minimal computational overhead. Extensive experiments on the DOTA, RSOD, and NWPU VHR-10 benchmarks demonstrate that ODANet achieves a superior balance between detection accuracy and efficiency compared with existing lightweight methods. In particular, it attains 77.8% mAP50 on DOTA at 42.3 FPS with only 48.3 GFLOPs. Furthermore, consistent performance on edge platforms validates its practicality for real-time remote sensing applications.