Front-End and Back-End Adaptations of Oriented R-CNN for Oriented Object Detection in Remote Sensing Images

Abstract

Oriented object detection in remote sensing images is challenged by arbitrary object orientations, large-scale variations, and complex backgrounds. In Oriented R-CNN, direct top-down feature fusion may attenuate local structural cues, while the shared representation in the RoI head may not adequately accommodate the different requirements of classification and rotated bounding box regression. To examine these issues, this paper introduces two targeted adaptations. A Feature-Aligned Aggregation Feature Pyramid Network (FAA-FPN) replaces direct cross-scale addition with difference-aware gated recalibration for multi-scale feature aggregation. A Decoupled Quality-Aware Rotated Detection Head (DQRH) constructs task-specific representations for classification and regression and incorporates quality-aware and angular auxiliary supervision during training. Experiments on DOTA-v1.0 show that the complete model achieves 76.88% mAP, which is 1.01 percentage points higher than the Oriented R-CNN baseline. Fine-grained ablation studies, loss-weight sensitivity analyses, and efficiency measurements further characterize the effects and computational costs of the proposed adaptations.