An Improved Oriented Object Detection Method for Remote Sensing Images Based on Oriented R-CNN

Oriented object detection refers to the process of using oriented bounding boxes to localize and identify objects. Oriented R-CNN is a classic oriented two-stage detection method, which is used as the benchmark model in many researches, but it still has some shortcomings in the following aspects: (1) Traditional convolutions' fixed geometric structure limits feature extraction for oriented objects, resulting in poor alignment with oriented bounding boxes. (2) The maximum IoU strategy can create a mismatch between classification and regression, affecting model performance. (3) A rigid detection head struggles to handle the complex geometry of oriented objects. To tackle the first issue, we introduce an alignment convolution to overcome the limitations of traditional convolution's rigid structure for oriented object feature extraction. For the second issue, we enhance the baseline model's sample matching strategy by incorporating task-aligned assignment to mitigate task misalignment. Finally, to address the third issue, we propose an improved detection head design to better handle complex object geometries. Ablation studies confirm the effectiveness of each proposed improvement and their combined impact. Our model outperforms leading oriented object detection methods on both the DOTA and DIOR-R datasets.