RA2Net: Rotated alignment and aggregation network for oriented object detection in aerial images

Min Dang, Qijie Xu, Gang Liu, Hao Li, Xu Wang

Abstract: Oriented object detection in aerial images presents unique challenges, including feature misalignment caused by rotation-agnostic convolutional backbones, severe scale and aspect ratio imbalances among densely packed objects, and instability in loss optimization due to angular discontinuity. Traditional horizontal detectors fail to capture orientation variance, while existing rotated detection methods often struggle with aligning rotated sampling points or applying uniform penalties regardless of object geometry. In this paper, we propose RA2Net (Rotated Alignment and Aggregation Network), a novel framework that dynamically aligns features and adaptively optimizes loss for robust oriented object detection. First, the Rotated Point Alignment (RPA) module addresses feature misalignment by predicting oriented bounding boxes (OBBs) and refining convolutional sampling points via rotated cross convolution, enabling rotation-equivariant feature extraction. Second, the Rotated Feature Aggregation (RFA) module integrates dual-branch attention to fuse multi-scale local and global features, selectively enhancing informative regions while suppressing background noise. Third, we introduce an Adaptive Geometric-aware Loss (AGL) that combines thermodynamic energy modeling for small objects with polar coordinate regression for elongated objects, thereby stabilizing training and improving localization accuracy. Extensive experiments on aerial object detection benchmarks demonstrate that RA2Net not only achieves high precision but also maintains robustness across diverse object scales and orientations.

External IDs:dblp:journals/ijon/DangXLLW26