Cross-Grid Label Assignment for Arbitrary-Oriented Object Detection in Aerial Images

As a challenging task in the field of remote sensing, object detection has attracted widespread attention from researchers. However, for aerial images with an imbalanced foreground-background distribution, the existing label assignment assigns insufficient positive samples to aerial objects, severely limiting detection performance. In this letter, we propose the cross-grid label assignment (CLA) to add high-quality positive samples used for training and loss calculation, thereby alleviating the issue of imbalanced positive and negative samples. Furthermore, the feature refinement head (FRHead), which extracts object-oriented features and guiding semantic enhancement, is used to address the inconsistent between classification scores and localization accuracy. Extensive experiments have shown that our method has superior detection performance, with 90.50% and 73.69% mAP on the HRSC2016 and DOTA datasets, respectively.